VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
# Electron parity ledger
Goal: all user-visible Tauri capabilities in the Electron app, using existing backend/domain code and the established T3-inspired shadcn/Base UI shell. A route or a screenshot alone does not count as parity.
Design contract
- Keep the expanding script editor, bottom generation controls and independent player.
- Use integrated, resizable panes for contextual work. Reserve dialogs for destructive confirmation.
- Keep each workspace's local controls in the shared readable secondary-sidebar hierarchy; it resizes on desktop, reflows above content when narrow and preserves a labelled rail when collapsed.
- Use existing brand assets, semantic theme tokens, local typography and Lucide icons.
- Simple primary flows; advanced controls remain available through progressive disclosure.
- Reuse domain/API behavior, not legacy Tauri CSS or embedded legacy pages.
- Preserve local-first behavior, existing user data and cross-platform capabilities.
- Every new visible string is translated into all 21 locales.
Capability inventory
| Area | Tauri source | Electron status |
|---|---|---|
| Voice cloning, reference upload/recording, playback | pages/CloneDesignTab.jsx, hooks/useTTS.js | Complete behavior parity: staged reference choice/upload/recording, local transcript recovery, script import/paste/cursor inserts, saved-voice switching, language and production controls, stable generation progress, cancellation, explicit completion feedback, paused completed-take Vidstack playback and reusable profile handoff are implemented; native verification remains recorded below |
| Voice design, direction, presets | pages/CloneDesignTab.jsx | Complete behavior parity: debounced local description mapping, conflict-safe category editing, personality/preset starting points, seed editing/reroll, language and production controls, stable synthesis progress and completion feedback, bundled demo playback, design profile save/edit/restore and complete recipe persistence are implemented; personality toggles and the collapsed recipe summary now match Tauri's current controls |
| Saved voices, profile editing and portraits | pages/VoiceProfile.jsx, hooks/useProfiles.js | Complete behavior parity: the full profile library has direct Voice cloning and Design creation paths, including an actionable empty state; clone/design editing, initials and uploaded/searched portraits, language/reference/style editing, persona export/import, usage and guarded project navigation, preview generation with manual playback, consent/revoke, identity lock/unlock and deletion are implemented in the shared resizable editor pane; native download/platform checks recorded below |
| Recent takes, star/reuse/export | components/WorkspaceHistory.jsx | Complete behavior parity: unified All/Clone/Design/Starred history, lazy Vidstack playback, mode-aware reopen/reuse, full take details and production recipe, native save, save-as-profile, identity lock, starring, deletion and clear-all implemented; A/B saved-voice/preset comparison remains available in Tools |
| Models and engine configuration | pages/ModelCatalogue.jsx | Complete behavior parity: selection, downloads/installers, searchable compatible catalogues, speaker-diarisation access, explicit preview consent/cache control, device-aware bundles and disk checks, reconnectable progress, interrupted-cache repair, installed/resident state, unload, confirmed delete and reinstall implemented; Windows model-cache directory selection/reset/cancellation are verified, while native cross-platform install verification remains recorded below |
| Pronunciation dictionary | components/settings/PronunciationPanel.jsx | Complete behavior parity: global and language-scoped respelling, IPA and CMU entries; live substitution preview; enable/delete; JSON add-or-replace backup restore; shared synthesis backend |
| Dictation and transcription | pages/Transcriptions.jsx, capture components | Complete behavior parity: installed Parakeet dictation, upload/record/live capture with the registered desktop shortcut shown in context, searchable history, cleaned/raw transcript detail, segment timings, copy/export/delete/clear, explicit script reuse and inline missing-model selection/install recovery are implemented; Electron additionally exposes input/channel controls and cancellable file transcription. Native Windows hold/toggle hotkeys and target insertion are verified; physical microphone and native macOS/Linux verification remain recorded below |
| Dubbing and timelines | pages/DubTab.jsx, hooks/useDubWorkflow.js | Complete behavior parity: file/URL upload, caption seeding with ASR fallback, SRT import, detailed streamed preparation/ASR state, persistent multi-language translation/generation, pasted translation mapping, glossary CRUD/auto-extraction, quality/reflection/condensing/dialect controls, strict LLM-ready Agent Fit with measured render/rewrite passes, failed-line retry/skip, source restore and segment cleanup, optional persisted stage-review checkpoints, complete segment/casting/timing editing with undo/redo and bulk actions, per-language changed-line regeneration, synchronized waveform/onset/QC tools, opt-in debounced live-as-you-edit CAST preview, manual previews and complete export controls are implemented; edit persistence and task reconnection are shared with Projects. A local one-speaker model-backed workflow is verified; multi-speaker verification remains below |
| Batch processing | pages/BatchQueue.jsx | Complete behavior parity: multi-file/multi-language enqueue, saved/default/designed/preset voice and background options, active/completed/failed views, live stage/language/segment progress, localized timestamps and completion duration, cancellation, confirmed deletion and per-language export; native watch folders provide stable-file ingest, pause/stop and streamed uploads to local or authenticated remote backends. A full local model-backed batch and server-resolved preset voice are verified |
| Stories | components/StoriesEditor.jsx | Complete behavior parity: character cast, per-line and selection-scoped voices, cursor-aware pauses and expressive delivery cues, ordering, chapters, global/per-line speed, local auto-cast, SRT/TXT/Markdown import, sentence splitting and an editable sample story; cancellable line previews, per-character WAV stems, streamed full rendering/stop/recovery, Vidstack playback/export, durable drafts and named projects are implemented. Full model/platform verification remains below |
| Audiobooks | pages/AudiobookTab.jsx | Complete behavior parity: script/file import and editable sample, cursor-aware pause/voice/pacing/emphasis/spelling/reaction markup, default and named cast voices, language, format, metadata/cover, loudness, pronunciation, statistics, chapter plan/previews, production overrides and durable projects; streamed rendering/stop, manifest recovery, synced Vidstack playback and native export are implemented. A real two-chapter M4B render/playback with chapter metadata is verified; broader platform/model verification remains below |
| Projects and recovery | pages/Projects.jsx, utils/projectState.js | Complete behavior parity: shared dubbing project open/save/save-copy/delete, legacy state preservation and reload recovery; unified searchable and filterable dubbing/Stories/Audiobook/profile/transcript/take/export/render library; card/list views, preview/reuse/reveal actions and inline rename implemented |
| Voice discovery/gallery | pages/VoiceGallery.jsx | Complete behavior parity: searchable/filterable archetypes, durable favorites, paginated Community catalogue, local and searched imports, inline trimming, Vidstack previews with local fallback, profile materialization, Designer/Stories/Audiobook handoffs, submissions and saved-clone persona export/import are implemented. External search/download, submissions and recorded-community verification remain below |
| Audio tools | pages/ToolsPage.jsx | Complete behavior parity: direction parser, speech-rate fit and media probe are implemented alongside Electron's voice comparison and full file/recorded voice-conversion flow with target selection, cancellation, recovery and inline missing-ASR setup. Media-tool install/status/system/restore and a custom executable picker are integrated. Model-backed conversion is verified with the selected CUDA ASR/TTS path; Windows native picker selection, authorization and cancellation are verified |
| First-run/setup and model readiness | pages/SetupWizard.jsx | Complete behavior parity: explicit managed-runtime install/location/retry, pre-runtime language and download-region choice, automatic system preflight, native permission visibility, invisible-until-needed media recovery, UI scale, required/curated model checklist, analytics choice, credentials/mirrors, inline dictation-model recovery with sample replay and automatic first-sound handoff are implemented in a sidebar-free setup shell. Uncached cross-platform installation remains in the verification ledger below |
| Full settings: devices, network, storage, privacy, logs | pages/Settings.jsx | Complete behavior parity: general, appearance, models, mirrors, backend/frontend logs; local-only usage insights; native microphone and macOS accessibility status/recovery; proxy and Hugging Face/DeepL/Microsoft credentials; compute-device override, compile workaround, generation budgets and hardware readouts; privacy/history retention, disk usage/warnings, temporary cleanup, database backup, scoped full-data reset, verified app-data relocation, secure model-cache relocation/reset, and native full-data uninstall; explicit LAN/Tailscale sharing and authenticated remote backend switching |
| Launchpad/support | pages/Launchpad.jsx, pages/SupportPage.jsx, pages/ContactPage.jsx | Complete behavior parity: responsive branded workspace entry, recent exports, direct profile/take/project continuation, unified library navigation, support, donation progress, sponsors, commercial licensing, contact/security/community channels, scrubbed diagnostic reports and desktop updates are implemented |
| Local repair agent | Electron extension | Codex, Claude Code, OpenCode and Pi are detected without a shell, with explicit diagnose/fix permissions, scrubbed app/log context, streaming output, cancellation, remembered opt-in automation and review-before-PR guidance. Explicit app setup/recovery runs work from packaged builds in an isolated temporary session through the scoped API bridge without demanding source; Electron restart and resumable setup controls remain available while Python is down, while clean reinstall remains an explicit user decision. Diagnosis and code repair remain gated to a writable checkout. The right-center launcher and content-pushing footer dock pass the native bridge smoke; this Windows host exposes installed Codex, Claude Code and OpenCode while correctly disabling unavailable Pi, and packaged Linux rejects Windows CLI shims injected by WSL PATH |
| Desktop lifecycle, files, update, tray, native shortcuts | src-tauri, utils/media.js | Complete behavior parity: single-instance lifecycle, trusted navigation, native files/capture/shortcuts, branded window/taskbar/tray assets, stable/preview desktop updates, channel-aware release history, renderer persistence flushing before window close, quit and update install, and bounded backend teardown are implemented; macOS/Linux packaging verification remains recorded below |
Current verification limits
- Localization catalogue coverage is complete: all English keys exist in all
20 translated catalogues (0 missing English-key/locale pairs). Run
node electron/tests/locale-coverage.mjs --detailsto verify. A live renderer smoke also switches through all 21 languages and covers representative long strings at 960 px and 640 px without horizontal document overflow. Human language review remains a separate gate. - All 21 locales pass encoding/interpolation checks. Startup/recovery and engine-readiness, recording, navigation, player and TTS warning namespaces, plus profile-save messages, have explicit coverage guards. Existing Tauri translations supplied 813 missing entries with matching English meaning and placeholders, including reviewed capitalization/punctuation differences in voice and recording labels.
- Latest live Windows check: backend and renderer are reachable; Max has four applicable engine families with OmniVoice, Faster-Whisper Turbo, Parakeet and NLLB selected. The Models UI switched Faster-Whisper between in-process and isolated runtimes and restored the original selection; real upload transcription, segment display, export, Clone handoff, persona export/import and cleanup passed against seven transcript-backed profiles, and System preflight rendered all eight checks. The complete audio.cpp Sortformer Q8 and the app-managed, checksummed audio.cpp v0.7.4 runtime are installed; pyannote remains incomplete. The managed runtime completed real multi-speaker inference and cancellation on the automatically selected RTX 4090 Vulkan device. No model download was started by this audit.
- Google may return a JavaScript-only image-search shell to non-browser clients. Profile search keeps Google as its first source and falls back to credential-free, sensitive-content-filtered Openverse thumbnails that allow modification and commercial use.
- Model install tests must mock downloads; user-triggered model downloads remain explicit.
- Windows development does not establish macOS/Linux packaged behavior.
- Backend installed-model guards must remain in place for automatic reference ASR.
- Large migrations require workflow regression coverage and real state restoration tests.
Open verification gates
| Gate | Current evidence | Needed proof |
|---|---|---|
| Localization | All 21 catalogues have complete key coverage, pass encoding/interpolation checks, and render through real language switching; representative long-string pages fit at 960 px and 640 px | Complete human language review |
| Clean first install | Setup shell, explicit runtime/model install, recovery, cached onboarding, and fully isolated packaged Windows and Ubuntu 26.04 dependency/model/onboarding/paused-first-sound runs with clean shutdown are exercised | Fully uncached packaged installer-to-first-sound run on macOS |
| Desktop platforms | Windows native flows plus launched AppImage/installed Debian Wayland shells, tray recorder, watch picker/lifecycle, clipboard delivery and 150% DPR layout are exercised | macOS native run, Linux portal shortcut/physical microphone, signed packages and updater runs |
| Physical audio | File-backed dictation, transcription, playback, and echo-cancellation routing are exercised | Physical microphone insertion and speaker-to-microphone AEC on each desktop OS |
| Diarisation | The installed Sortformer Q8 model plus the managed audio.cpp v0.7.4 Vulkan runtime produced four attributed turns across three detected speakers from a 12-second real two-person clip; cancellation stopped the owned process in 1.17 seconds with no registered process left. A 35-second run completed through the production adapter with exact media bounds, and the full supported 119.5-second v1 window completed in 6.5 seconds with 40 bounded turns/four detected speakers. | Long-recording production run through pyannote or locally converted Sortformer v2.1 streaming |
| Authenticated media imports | Cookies-file and caption fallback paths are implemented; a live public YouTube ingest downloaded playable media plus one authoritative original-language caption track (165 parsed cues) without loading ASR | Live authenticated URL and caption download owned by the user |
Maintenance
Renderer-side updater, dictation setup, locale-chunk, translation-file and local-agent discovery failures are contained at their initiating UI boundary. The live 37-route sweep reports zero exceptions, warnings or failed API requests; real transcription, persona round-trip and five-result portrait persistence also pass against the running backend.
Caption-first URL ingest now detects yt-dlp's explicit
*-origautomatic-caption tracks when a source omits language metadata. Runuv run --directory backend python ../scripts/smoke_dub_url_captions.pyfor a live, model-free download/parse check that cleans its isolated media directory.Sortformer's managed-runtime action now appears only after the status contract confirms platform support. Status failures remain visible with an inline retry; a focused query-state regression covers failure, hidden installation, recovery and the supported install action.
Uploaded-reference transcription failures now open Optional details, state that transcription failed, and keep Retry transcription beside the editable transcript. Profile saving remains available so a user can supply text manually or retain the backend's best-effort recovery path; the focused profile-save regression covers the failed state and retry action.
The shared TTS status contract no longer publishes a stale
readysub-stage after model unload. Electron therefore receives one coherent idle/not-resident state while genuine load failures keep their diagnostic detail; the focused backend status suite covers both transitions.The Electron skeleton now resolves against the current published dependency set:
cn0.3, PostHog 1.430.3, JSDOM 30.0.1, TanStack Router 1.170.36 and React i18next 17.0.14, with Bun 1.4.2 recorded as the workspace package manager.bun outdated --latestreports only Node 26 typings; the project intentionally stays on the current Node 24 typings that match Electron 44's runtime. Electron typecheck and the focused renderer checks pass. The production renderer bundle and native repair-agent launcher smoke also pass with the updated runtime graph.Collapsed Engine Ready no longer imposes a workspace-navigation scroll at the standard 720 px window height: all nine destinations remain visible. Expanded engine details and windows at 680 px or shorter retain a bounded 32vh list so the engine controls and Settings footer stay reachable. The responsive smoke now waits for tooltip dismissal before screenshots and asserts the normal-height navigation has no hidden rows. The route-shell regression now follows the current mechanics too: one 48 px rail remains mounted across all nine secondary workspaces without renderer errors or overflow, then Clone restores the full saved-voice library automatically.
Home now uses the same TanStack lazy route boundary as every feature workspace, and longform/project persistence modules load through the asynchronous native close hook instead of the root renderer graph. The minified production entry fell from 1.55 MB to 1.42 MB (8.7%) while keeping persistence flush behavior; the production build passes and a live Home render has no page error or horizontal overflow.
Constrained Dubbing and secondary-pane routes now switch the primary library to its existing 48 px navigation rail through 1680 px, including Windows display scaling, while a manual expansion remains available for the current route. The responsive smoke covers every secondary workspace at 1920, 1440 and 960 px, guarantees at least a 480 px Dubbing canvas, and verifies the stacked Dubbing layout plus full-width Settings content at 640 px. All 21 locale layouts and representative long strings pass without document overflow.
Electron locale coverage is complete across all 21 catalogues: 3,059 missing English-key/locale pairs were filled from established Tauri translations, guarded machine translation and manual repair. Encoding and interpolation validation passes with zero missing keys. The Google translation backend now rejects generic HTTP-200 error-page text and replacement characters instead of overwriting transcript rows; focused retry/fallback regressions pass.
NLLB silent fallback fixed in source: expanded short aliases, tokenizer-supported FLORES and unambiguous ISO-639-3 codes; unsupported source/target/segment codes return a structured 400 before inference. Electron maps that response to actionable copy in all 21 locales. Eight backend offline/empty-cache checks, two renderer error-mapping checks and typecheck pass.
Interactive and queued Dubbing now share hardware-bounded native OmniVoice batches, reducing N renderable subtitles to ceil(N/W) model forwards on capable GPUs while retaining serial/OOM fallback and fixed-seed determinism. NLLB groups rows by target language into bounded forward passes, preserves request order and retries a failed group per row. Sixty-six focused Dubbing, incremental-render, Smart Fit and translation checks pass offline.
Long Dubbing transcripts now coalesce each short typing burst into one undo snapshot instead of cloning the complete segment history per character. Offscreen segment rows use browser content visibility containment, reducing layout and paint work without changing the existing timeline, selection or editor mechanics. The fail-before undo regression and all fourteen Dubbing-session checks pass; Electron typecheck and targeted lint are clean.
Completed Dubbing transcription is now marked in the durable job before its final SSE event. A reconnect or managed-backend restart replays the saved segments, language and cast immediately instead of launching a duplicate whole-file ASR pass; the fail-before reconnect regression and all active transcription-stream checks pass offline (#2051). A live Windows run transcribed a 16.98-second clip into three segments in 4.75 seconds, restarted the managed backend, then replayed the identical result in 24 ms with no ASR start or intermediate-segment event.
Saved Voices now applies the same large-list windowing as Recent Takes while preserving the full responsive Profiles grid. A live 100-profile browser smoke mounted 10 sidebar rows, scrolled through the shared pane, and reached the final profile; small libraries retain their established compact row mechanics.
TTS readiness now asks for the exact operation before enabling work across Clone, Design, Stories, Audiobook, Dubbing and Batch. Clone filters readiness to engines that advertise cloning instead of allowing a fixed voice to fail after click.
Performance Settings now uses explicit Fast/Balanced/Quality/Max choices for the global profile and every engine family. The visible Balanced default and every later tier select only already-installed, locale-compatible ASR/Dictation models, while explicit environment pins and ready network translation providers remain authoritative. Persisted profiles reconcile at startup, preventing a missing historical selection or unavailable network translator from hiding compatible local weights; unsupported families stay disabled and no tier starts a download. The live Windows backend reports OmniVoice, Faster-Whisper Turbo, Parakeet and ready local NLLB selections; a 1280 px browser pass rendered all seven controls without overflow or page errors. The model-aware settings smoke, Electron typecheck, targeted lint and ten offline backend checks pass.
Every engine/model picker now confirms the accepted selection. Picks routed through
/engines/select, including ASR choices made from the model catalogue, consume the backend verdict: CPU fallbacks warn with their reason, while accelerated selections surface VRAM or driver caveats before the first generation. Dictation, Translation and Diarisation receive the same clear completion feedback. Helper and component regressions cover fallback, accelerated-caveat and ordinary results; all 21 locale catalogues include the reused Tauri copy.Fixed initial-language rendering: React now subscribes to resource-store additions, so a saved language bundle replaces English fallback labels as soon as it loads. Rendered regression failed before the fix; all 14 focused language checks pass afterward.
Profile save now enforces transcription/pending guards in the submit handler as well as the disabled button, preventing alternate form submission from saving before the reference transcript is ready. Typecheck passes. Creation already sends audio and portrait together in one request, avoiding a client-side second-upload partial-save sequence.
Focused rendered-form regression confirms direct submission makes no save request during transcription, retains the draft, and saves successfully after transcription completes.
Language switching now preserves the latest user choice if an earlier translation bundle loads later. Regression reproduced an older German request replacing a newer English choice; it now passes and checks UI language, document language and persisted preference.
Desktop/browser locale selection now understands Chinese script tags (
zh-Hant,zh-Hans), region tags, underscore-separated OS locales and Unicode extensions. Explicit script preference wins over region, preventing Traditional Chinese users from receiving Simplified Chinese. Twelve focused cases and typecheck pass.Reference transcript races verified: replacement uploads abort the prior request and ignore its late result; selecting a saved profile protects even its empty transcript. Six focused hook checks pass. Missing-model UI coverage checks the backend response, not actual inference prevention.
Global performance changes no longer replace explicit Dubbing production steps. Unset steps use backend presets; targeted hook regression and typecheck pass.
Keep this ledger current with each migrated workflow. Run focused regression tests while editing, then typecheck/lint/build. Recheck API contracts against the running backend after restarting it. Preserve native button/link semantics and existing composer geometry. Do not increment the application version without owner direction. The parity-inventory regression classifies every command in Tauri's native handler by one of the areas above, so a newly registered shell capability fails CI until Electron receives an explicit behavior-parity decision.
Verified checkpoint
The desktop updater now has direct state-machine coverage for Stable and Preview platform feeds, available/downloaded progress, quiet startup failure, manual error reporting, guarded download/install and trusted renderer IPC.
Electron safely renders available-update and GitHub release Markdown through the shared Tauri tokenizer, preserving headings, bullets, bold text and code while leaving links and raw HTML inert. Release history therefore matches Tauri without adding an HTML injection path.
Persistent Electron update errors expose the same Retry and Dismiss recovery as Tauri; dismiss clears transient update metadata through trusted main-process IPC without changing the channel.
Available or downloaded Electron updates remain visible in the existing sidebar notification center after the one-time toast, including while Python is unavailable. The indicator opens Updates directly and stays until the updater reports a newer state.
The 2026-09-13 Windows parity recheck passed 37-route DOM accessibility at 1440 and 640 px, all nine secondary-sidebar layouts from 1920 through 640 px, live CUDA engine/readiness disclosure, reversible Faster-Whisper runtime switching, real Faster-Whisper upload transcription/export/Clone handoff, and persona export/import with and without its reference clip. The current inventory remains seven transcript-backed profiles with TTS, ASR, translation, dictation and diarisation ready; disposable transcription and persona records were cleaned.
The collapsed launchpad keeps the branded app rail and replaces the header brand mark beside
VoiceStudiowith the explicit sidebar-expand control. A rendered 900 px browser check found the localizedToggle Sidebarbutton in that position with no duplicate header image.A repeatable read-only live-state smoke now cross-checks backend selection, installed inventory and readiness against the rendered Engine Ready panel, all five applicable model-family pages, Transcribe and Saved Voices. The current Windows run reports ready CUDA TTS/ASR, NLLB, Parakeet and Sortformer, complete setup, hardware details, and no renderer/backend errors.
The live saved-profile library currently contains seven profiles and every profile has a persisted reference transcript; the upload, installed-ASR and saved-profile handoff has no outstanding transcript recovery item in current user data.
The current sidebar verifier distinguishes the six visible engine actions from the six hidden expanded-detail links and removes its screenshot directory after each successful or failed run. Route layout covers all 38 rendered routes at 1440 px and 640 px without overflow.
Blank-transcript profile saves now reuse the strongest compatible ASR model already installed on disk when the selected ASR and Dictation models are stale or missing. A live Windows save transcribed an existing reference with Faster-Whisper Large-v3 Turbo in 2.88 s, preserved both engine selections, left the installed-model inventory unchanged, and cleaned the disposable profile. The fallback receives a concrete local snapshot path, so it cannot trigger an implicit Hugging Face download.
Startup now reconciles the visible Balanced performance default with the installed inventory. A live Windows restart replaced stale missing Whisper Large-v3 and Dictation Tiny selections with installed Faster-Whisper Turbo and Parakeet TDT v3; both readiness endpoints became ready, Transcribe lost its false missing-model panel, all four applicable engine families appeared, and the install queue remained empty.
The shared Fast/Balanced/Quality/Max TTS defaults now cover Dubbing, Batch and Voice Conversion as well as Clone and long-form rendering. Untouched Dubbing requests omit
num_step, explicit Production overrides still win, and the live Quality contract resolves OmniVoice to 32 steps without starting a download.First-install Models now leads with the existing hardware-specific recommendation summary and keeps only required or already-installed models expanded. The full compatible catalogue remains available in one disclosure, reducing the primary CUDA setup view from many model cards to the four that describe this machine's current state. A disposable 960 px browser session verified all four setup gates without changing onboarding or model data.
Electron marks deliberate Windows backend shutdowns before terminating the process tree, so normal quits and restarts are not reported as crashes; health-supervision kills remain crash-reportable.
Electron now installs explicit Chromium permission check and request handlers before any window loads. The trusted app and development origins receive audio-only microphone access plus the clipboard/fullscreen permissions used by existing controls; camera requests, foreign frames and every unrelated permission remain denied. This matches Tauri's reliable WebView microphone grant without broadening the renderer's device access; native physical-microphone verification remains a platform release gate.
Electron now enforces Tauri's whole-window render invariant from the main process: an empty or unreachable renderer gets three bounded app-URL reloads, continued low-frequency health probes, and then a self-contained localized retry page that does not depend on React or renderer assets. Retry clears only Chromium's display cache and relaunches through the orderly backend-shutdown path, matching Tauri's corrupted-WebView recovery without touching app data.
Once a fresh install enters onboarding, that requirement survives a restart after the required model finishes installing; privacy and dictation can no longer be skipped merely because model readiness changed between launches.
Live keyless portrait search now returns five normalized in-app choices when Google serves its JavaScript-only shell, using Openverse's non-sensitive, modification- and commercial-use-compatible fallback results.
The ASR model page now addresses the OpenAI-compatible configuration routes through the settings API namespace, removing its false load-error banner and restoring save/test actions.
Live installed-model verification now covers both transcription tiers against a saved 16.98 s profile reference: Parakeet TDT v3 produced the complete transcript in 2.08 s through Fast mode, and the selected CUDA Faster-Whisper Large-v3 Turbo produced the same transcript with three timed segments in 2.68 s through Accurate mode. The profile portrait and canonical audio endpoints both return their correct media types and complete bytes; no model download or profile mutation was performed.
A live clone generation using the saved Scarlett profile and the exact script
Native delivery works.completed through CUDA OmniVoice in 0.85 s. Accurate Faster-Whisper recognized onlyNative delivery works.from the generated audio, with no reference bleed or leading words. The resulting TTS residency reports the activek2-fsa/OmniVoicecheckpoint oncuda:0with measured VRAM. Loaded dictation inventory now derives its device from the backend contract and uses the selected model label, preventing CPU-only Parakeet from being presented as CUDA after the next backend start.Electron's current dependency baseline is up to date apart from the intentionally held jsdom 30 major: React/Base UI/shadcn/TanStack/Vite/Electron remain current, while React Virtual 3.14.12, HLS.js 1.7.3, WaveSurfer 7.12.12 and React 19.3 types are now on their latest compatible releases. Typecheck, source lint and the production build pass after the update.
Window close, application quit and update installation now request a bounded renderer persistence flush before native teardown. Working Stories/Audiobook drafts drain synchronously and queued IndexedDB project mutations settle before acknowledgement; older live preload bridges remain safe during development HMR.
Electron startup and its managed backend were exercised together on Windows: the ESM main bundle now loads the CommonJS updater safely, the supervisor reaches
ready, and the core model/profile/history/translation/dictation routes return 200 through the renderer proxy. Dictation warmup no longer overwrites the TTS model status;/model/statusremainsModel readyacross the delayed ASR preload.Vidstack owns all Electron playback; WaveSurfer renders waveforms only. The shared provider selects native audio/video, HLS, DASH, YouTube, and Vimeo sources, lazy-loading the streaming libraries. Gallery search results use the YouTube provider for in-place preview.
A current source audit confirms every Electron play/pause path still reaches
StudioMediaPlayer; no feature renders native audio/video controls or asks WaveSurfer to play. The solenew Audio()use is the metadata-only duration probe, which never starts playback.Browser regression: manual play, pause, actual seek position, exclusive profile preview, and output dismissal (
node electron/tests/playback-smoke.mjs). No user data is changed.Explicit model download/cancel and sidecar installation controls; reconnectable download status.
Installed models now expose live memory residency, safe unload, confirmed delete and confirmed reinstall. Interrupted caches show their actual disk use and offer one-click repair or removal. Removal preserves FastAPI path routing for slash-separated Hugging Face repository IDs; a resident model is unloaded before its files are removed. Focused UI regressions cover operation order and the delete/install/repair contracts. Actual model files are not changed by these tests.
Model settings now keeps the speaker-diarisation catalogue reachable after onboarding. Every family has metadata search, while incompatible models stay folded until explicitly requested. Focused regressions cover diarisation routing, metadata search and compatibility disclosure.
Installed CTranslate2 ASR weights can now be selected as the active Faster-Whisper model. The selection persists, updates engine status, and Dubbing turns typed missing-ASR failures into an installed-model recovery action instead of exposing
asr_model_missing.A root-level model-install observer follows jobs across route changes, then refreshes catalogue, engine availability, setup readiness, residency and dictation state as one completion transition. Successfully installed ASR/TTS models become selectable immediately, while delete and reinstall refresh the same dependent views. Stable engines stop background catalogue/engine polling; generation publishes its faster live model status directly to Engine Ready and refreshes residency when it finishes.
Selected historical takes now expose the same native Save As flow as Latest take, including reveal-on-success and shared failure handling. A focused regression covers the native request.
Saving an uploaded reference as a profile now atomically releases the temporary file and selects the returned profile with its canonical language, style and local-ASR transcript. This prevents a deleted or cleared profile from falling back to stale pre-save audio, and the save action names reference transcription while it waits instead of appearing inexplicably disabled.
Translation family selection persists locally; LLM provider credentials, probes and model listing are implemented; per-skill routing is implemented; non-LLM DeepL/Microsoft credentials implemented in Settings > Credentials. Installed Python dependencies are not proof of downloaded weights.
Renderer test suite, typecheck, lint, production build and frozen Electron lockfile passed at their recorded checkpoints; rerun affected checks before landing.
Four backend regressions cover download status and installed-only translation selection.
YouTube is connected to a migrated workflow. Vimeo, HLS and DASH have working shared provider selection but still need live sources from their eventual migrated workflows for end-to-end verification.
Voice-design requests verified in-browser; clone references are excluded. Tauri typecheck and 26 shared transcription/seed/attribute tests pass after helper reuse.
Dubbing browser smoke passes with synthetic audio and MP4 playback. Backend task responses are mocked; real dubbing/export and recovery are not yet verified.
Settings Logs browser smoke verifies backend filtering, explicit clear confirmation, and frontend console capture. The existing Tauri console buffer is reused.
Dropped dubbing task streams now preserve Resume/Cancel recovery and block duplicate jobs and edits until resolved. Regression tests distinguish disconnects from terminal errors.
Dubbing SRT/URL import browser checks pass with mocked backend responses; 19 offline backend subtitle parser/voice-metadata tests pass. The isolated live downloader also produced a browser-safe MP4 and parsed 165 original-language caption cues.
Gallery browser smoke covers catalogue browsing, click-only preview, profile handoff and navigation during saving. Live catalogue read returned 1,126 entries / 7 categories.
Gallery favorites persist across reloads and include matching voices from later pages; browser and unit regressions cover both behaviors and facet query parameters.
Native MP4 checks verify mute, volume, and fullscreen enter/exit. Gallery missing-audio responses produce a visible unavailable control; exclusive audio playback remains verified.
Batch browser smoke verifies enqueue, live progress display, renderer reload without resubmission, cancellation and confirmed deletion. A real local model-backed batch now verifies Turbo ASR, NLLB, saved-profile TTS, muxing and deletion.
Batch voice choices now resolve at the shared backend boundary: clone profiles supply their reference, designed profiles supply their healed instruction and seed, and legacy Tauri presets supply their actual design instruction instead of silently using the default voice. Invalid or deleted choices fail before a large upload is saved. A live
preset:narratorjob completed extraction, CUDA Turbo ASR, translation, CUDA OmniVoice generation and MP4 muxing; its input, output and queue record were deleted after inspection.LLM provider browser smoke covers environment pinning, blank-key preservation, save-before-probe, failed-save guard, model fetching and explicit activation.
Per-skill LLM routing browser checks cover independent disable, configured provider selection and preservation of unavailable stored overrides; the live backend lists six skills.
Dictation cleanup browser regression verifies opt-in settings, disabled subordinate controls, outgoing refinement flag and separate preservation of raw/refined text.
Tools browser smoke verifies direction/rate/probe requests and results. Live probe initially reported missing ffprobe (501); the app-managed installer resolved it and a real MP4 probe now passes.
App-managed media bundle installation completed; real ffprobe read 2 streams / 4 s. Media-tool browser smoke verifies explicit acquisition, progress locking and system selection.
Project browser smoke verifies legacy open/cancel, save, preserved settings, reload, and confirmed deletion. Real backend create/read/update/delete round trip passed with a disposable fixture.
Longform smoke verifies Stories/Audiobook requests, voice guards, shared pause compilation, independent Vidstack playback and reload persistence. Unit tests cover interrupted streams, explicit manifest resume, cancellation and duplicate-render prevention. Single-chapter real verification is recorded below.
Named longform projects use the existing IndexedDB adapter under an Electron-specific database. Browser smoke verifies actual durable save/reload/open/delete; unit tests cover overlapping saves and failed writes. Working draft storage and Tauri data migration remain separate concerns.
Live audiobook check passed with the installed OmniVoice model and a saved clone profile: one English chapter, cache reuse, MP3/M4B output and Electron playback. This does not establish other engines or multi-chapter recovery. Shared JSON request headers are now enforced (except explicit overrides and multipart), fixing real backend rejection that permissive mocks missed.
Stories line auditions and per-character WAV exports pass browser playback/download checks. The assembler now shares canonical voice/pause/speed parsing with full renders; arbitrary character IDs cannot collide with object prototype properties.
Unified project list and rename pass browser checks for both storage backends; a real dubbing rename preserves saved state. Local rename serializes with saves and cannot restore a deleted project.
Projects now matches the shared all-content library: stable counted filters cover productions, profiles, transcripts, takes and exports; rows provide direct voice/script reuse, Vidstack preview and native reveal actions. Projects joins the responsive secondary-sidebar routes, so the main library becomes its compact rail at narrow widths. Browser regressions cover filter counts, source visibility and existing project lifecycle behavior.
Connection settings browser checks verify proxy persistence/failure handling, local-only token reads, explicit validation, secret-input clearing and separately opted-in CLI token deletion. Real token-state schema verified without changing credentials. Shared Tauri Network/Translation tests pass.
Compute/performance browser checks cover saved versus active state, failed writes, environment pins, platform guards and timeout budgets. Live read-only API shapes pass; native cross-platform execution remains unverified.
Privacy browser checks cover explicit analytics consent, watermark control, failed saves, unavailable capabilities and confirmed retention reductions. Shared Tauri privacy tests and live read-only contract checks pass; Electron does not add a frontend analytics SDK.
Storage browser checks verify partial scans, folder reveal, confirmed cleanup and partial failures. Live reports and backup state were read successfully without deleting files. Shared Tauri storage tests pass; storage labels/messages are translated in all 21 locales.
Dubbing Fast/Autofit/Cinematic selection, draft/project persistence and backend-reported Fast fallback pass browser and unit checks. Missing-LLM guidance links directly to configuration; real LLM-backed quality output remains unverified.
Inline dubbing export format/track/background/subtitle controls and language-specific stems/clips added. Browser fixtures verify MP3/SRT downloads and retry; request-builder and native filename-filter tests pass. A real single-language MP4 export and server-resolved Batch preset path are verified.
Dubbing timing and voice-match choices now reach generation and persist through drafts/projects; submitted/completed timing is retained separately for recovery and export guards. Browser and unit checks pass; real multi-speaker timing verification remains pending.
Interrupted ASR now retains explicit Retry/Cancel across reloads and reuses prepared media. Browser and unit regressions confirm no automatic restart or duplicate upload; real mid-ASR interruption remains unverified.
Spoken-language and speaker-count hints now cover local/URL imports and persist through drafts and ASR retry. Browser and request-contract tests pass; live public URL ingest is verified.
URL imports accept an explicit per-import cookies.txt export, with shared Tauri limits/transport checks and no draft persistence. Browser and Tauri regressions pass; authenticated live downloads remain unverified.
URL caption-download opt-in now seeds the editor from the closest source-language track and skips ASR when usable cues exist. Missing or malformed tracks fall back to ASR automatically; a live public download produced 165 usable cues.
Dubbing production overrides now expose persisted steps/guidance/speed/global direction and reset. Browser request tests cover zero guidance; engine-specific audible effects and Smart Fit tuning remain pending.
Saved Smart Fit preferences now affect Electron generation when Smart Fit is selected; project round-trip and strategy-specific request tests pass. Real fitted output remains unverified.
Real Smart Fit export integration now verified using app-installed FFmpeg/ffprobe: 43 tests pass, including seven previously skipped real codec/endpoint cases. End-to-end Electron multi-speaker dubbing remains unverified.
Export choices now persist and populate from legacy project preferences; browser reload and project compatibility tests pass.
Segment volume/speed/direction controls added; dropped gain forwarding fixed, including zero gain. Browser reload/request checks and session tests pass.
The Electron dubbing timeline now renders the extracted source waveform, follows the shared Vidstack preview clock with a live playhead/click-to-seek, consumes the existing cached speech-onset analysis, and snaps move/resize gestures to speech, the playhead, adjacent edges and the low-zoom time grid. Selected slots play their source range directly and can synthesize a separate fast line audition through Vidstack using the same resolved voice, style, language, speed and strict-slot duration as generation. Its roving keyboard editor matches Tauri: arrows select, Enter toggles precise editing, Shift adjusts the end, Alt moves, S snaps to speech and Delete selects a neighbour. Exact time fields commit once through the same clamped move/resize path, preserving attribution and one-step undo. Alt-drag preserves controlled overlaps.
Finished dubs now expose Tauri's opt-in second-pass ASR verification. Results annotate only the matching unchanged segment revision, switch the preview to the checked dubbed track, surface flagged lines with the recognized text, and are invalidated when an output-affecting segment field changes. The verification remains non-destructive to script text and generated media.
Dubbing now keeps additional target languages in the durable draft/project state and renders each language sequentially through the existing translate/generate contracts. Stored complete translations are reused; incomplete translations are skipped instead of producing a wrong-language track. The editor restores the primary language after the batch, track preview/export retain all backend tracks, cancellation stops the outer batch, and the footer reports the active language and batch position. Editing source text invalidates derived translations while editing a translated line updates only that language.
Hugging Face mirror settings now cover Auto/preset/custom selection and explicit endpoint testing. Browser checks and live cached-state schema pass; all new mirror strings cover 21 locales.
Transcription recording now shares input-device/channel controls with cloning; browser selection checks pass. Native global dictation capture/insertion remains pending.
Shared recorder navigation races fixed: late permission is released, cleanup is aborted and closed views receive no recording callbacks. Both regression tests failed before the fix and pass afterward.
Transcription exports now retain date/language metadata and segment labels show both known bounds via shared Tauri formatters. Shared page tests and browser file-content checks pass.
Transcription history gains confirmed Clear all; browser checks cover cancellation and storage failure without losing entries. Individual deletion preserves newer stored entries.
Windows unpacked packaging succeeds. Resource audit found and fixed omitted Python-project README/license; packaging-contract check added. Packaged startup now passes: fixed a duplicate
requiredeclaration introduced by the ESM/CommonJS build shim. The Windows artifact opens its renderer, exposes the preload bridge, loads model settings and connects to the existing live backend. Distribution commands now check built main/preload syntax before packaging. Clean-machine bootstrap still requires validation (current supervisor expects uv or an existing venv).Packaged runtime now waits for an explicit install action, stages bundled code into userData/runtime/project, installs Python/dependencies with frozen uv, records successful dependency state, and supports cancel/retry. Seven runtime tests and native Windows first-run/no-install smoke pass; normal live-backend attach still passes. Actual multi-GB clean-machine installation, disk/path preflight, custom/portable paths, Tauri environment reuse and mirror recovery remain unverified/incomplete.
Runtime setup now checks free space/write access, reuses its private uv on retry and reports real setup phases. Windows fresh-environment frozen install/import and packaged explicit-install-to-managed-backend checks pass (dependency cache reused; model downloads disabled). Packaged Electron also discovers default, custom and recorded portable Tauri environment roots and reuses one only when its interpreter is structurally intact and both frozen dependency manifests exactly match the Electron bundle. First run shows the destination before download, supports a native folder picker and restores the default location in one click. Language and persisted download-region choices are available before Python exists; Auto probes the direct and restricted paths, while region presets apply the same resilient uv/Python/PyPI/Hugging Face routes as Tauri. Custom runtime ownership is recorded only when Electron installs into it, so uninstall includes Electron-created environments while preserving compatible Tauri environments it merely reuses. Cross-platform/uncached installation remains pending.
Installed-runtime relaunch also passes with a managed backend. Setup storage failures now carry localized remediation; native dictation requires the Tauri shortcut (including hold/Wayland) and session-bound output delivery contracts, not only an Electron global shortcut registration.
Native backend status now pauses all renderer queries before readiness and during restart/crash, then resumes active queries on
ready. A regression covers attach, ready and crash transitions; this removes the predictable 502/503 request fan-out without hiding recovery controls or logs.Native dictation delivery now has a shared Rust stdio helper compiling the Tauri implementation directly, plus an Electron transport with no delivery replay. Windows helper build/19 tests, Tauri delivery/16 tests, three transport tests and a real non-writing protocol smoke pass. Shortcut/capture integration, packaging and actual target-app insertion remain pending.
Native delivery helper is now built for the installer architecture and bundled outside ASAR. Windows artifact contract, packaged helper protocol and launch from Electron itself pass. Global shortcut/capture integration and macOS/Linux native-runner verification remain pending.
Fixed a prerequisite for native live dictation: the shared backend origin policy now includes exactly
app://voicestudio, with the same desktop defaults used by CORS and WebSocket checks. 64 origin/stream tests pass, including real endpoint handshake/final frames (stub ASR), lookalike rejection and explicit allowlist override. Capture UI/native shortcut wiring remains pending.In-view live dictation now uses installed-model preflight, shared PCM capture, real live text, pause/stop/cancel and shared history. Eight controller tests, browser smoke and real installed Parakeet with prerecorded Chromium input pass. AudioWorklet initialization cleanup is fixed in the shared Tauri helper (fail-before/pass-after). Production worklet hash matches source; typecheck/lint/build and 528 locale checks pass. Native global shortcut/widget/output integration remains pending.
Tray dictation now opens a dedicated recorder and integrates the shared native session/output helper. Eleven session regressions cover early releases, repeat presses, stale sessions, delivery ordering, clipboard fallback and no replay after uncertain output. Windows native smoke passes lifecycle, history and renderer IPC isolation with intercepted output; real target-app insertion, global hold/toggle/Wayland shortcuts and non-Windows runs remain pending. Closing the main window retains the prior quit behavior; macOS activation recreates the main window even with a hidden recorder.
Real Windows cross-app insertion now passes with two ordered utterances and clipboard restoration in a disposable target process (fake microphone/deterministic ASR). This caught and fixed a shared Tauri/Electron paste-consumption race: keep native delivery serialized through the existing 300 ms consumption interval before another utterance changes clipboard/modifier state. Fail-before/pass-after native smoke passes twice; 19 helper and 16 Tauri output tests pass. Global shortcuts, actual microphone input and native macOS/Linux verification remain pending.
Native shortcut backend now builds on Windows, Linux and macOS ARM (cross-checks for Linux/macOS). Windows real Ctrl+Alt+F24 registration/press/release and rollback smoke pass. The helper captures the destination at key-down and transports its session ID; Electron accepts unsolicited shortcut events separately from RPC replies (four transport tests). Wayland portal implementation/tests are shared with Tauri via a shell-independent core and separate Electron desktop identity. Global settings, hold/toggle dispatch and recorder session adoption remain pending; native Linux/macOS interaction is not verified.
Global shortcut settings now drive native capture with atomic accelerator persistence, shared backend hold/toggle mode, disabled-state unregister and Tauri's shared key recorder. Native key-down session adoption avoids recapturing focus; early key-up is retained. Windows native UI/shortcut smoke passes recording/save, saved file, hold, toggle and disable. 19 session/settings and seven Tauri hotkey tests pass. All 21 locales populated; one existing missing-key gap fixed in 18 root locales. Native non-Windows runs and capture restart during a previous transcription remain incomplete/unverified.
Corrected the earlier restart gap assessment against Tauri source: Tauri intentionally rejects new capture while transcribing. Electron now matches restartable done/error states without requiring dismissal. Session-tagged cancellation cannot cancel a replacement; an early hold release survives cleanup. Native restart smoke failed before/pass after; 17 session tests pass. Fixed native development's WebSocket routing (failed before/pass after) and verified real installed Parakeet through Electron+dev proxy with prerecorded microphone input; 10 live-dictation tests pass. Physical microphones and native non-Windows interaction remain unverified.
Batch watch folders now use the same pinned-directory Rust core and pure ingest tracker as Tauri. Native selection scopes access to a session capability; scans skip old files, new videos settle across two polls, and uploads stream directly to the trusted local backend. Pause/stop, current language/voice/background options, teardown/reload revocation and all 21 locales are wired. A separate helper isolates uploads from dictation. Windows native protocol/upload and Electron UI smokes pass; 8 shared Rust, 33 existing Tauri watch and 12 Electron controller/IPC/transport tests pass. Linux/macOS compile checks pass; native dialogs and complete model-backed batch execution remain unverified there.
Gallery My Imports now supports local audio/video uploads, explicit URL/search downloads, portable
.ovsvoice/legacy.omnivoicebundle import, Vidstack preview, confirmed deletion, the shared 60-second waveform trimmer and saved-profile cloning handoff. Shared API response declarations remain the source of truth. Real local upload, playback and profile handoff pass against the running backend with disposable fixtures cleaned up. Request/UI tests and Electron lint/typecheck/build plus Tauri typecheck pass. External search/download and bundle contents are not yet end-to-end verified.Archetype and imported gallery voices now offer Base UI destination menus for Stories and the Audiobook default. Handoffs preserve scripts, titles, existing cast and speaker mappings; repeated use of the same profile does not duplicate the Stories cast. Production disables the actions and is guarded again on mutation. Three state regressions and real-backend browser handoffs pass for both destinations, alongside the existing import tests and build/typecheck/lint. Imported clips use the shared 60-second waveform trimmer before these handoffs.
Community gallery now loads on explicit tab entry and supports paginated catalogue, source-scoped favorites, Vidstack previews, direct Designer handoff, profile materialization and clone/Stories/Audiobook destinations. Submission actions open the backend-generated form only on click. Shared response types are extracted from the Tauri API. Real catalogue (24 presets), favorite, preview and Designer handoff pass. The recorded-voice backend regression verifies idempotent clone materialization and source-audio refresh, while the Electron interaction regression verifies clone handoff. The configured live catalogue currently contains no recorded item, so that live materialization run remains unavailable. External submissions were not sent.
Shared Vidstack preview buttons now retry malformed archetype/community-preset audio once through the backend's explicit local-preview URL, matching Tauri. Browser regression fails before/pass after; a second fixture verifies a failed fallback stops without another request. Non-preset previews retain their existing behavior. Both routes still load only after the user starts preview. Portable persona export/inspection is implemented through the shared saved-profile editor.
Imported clips can now be trimmed inline with draggable waveform selection, numeric bounds, zoom/fit controls, looping Vidstack playback and keyboard play/save/cancel. Shared Tauri decode/slice/WAV utilities produce both the selection preview and saved crop; original audio is retained and source metadata copied. Bounds enforce the same 15-second reference limit. Real backend smoke verifies the visible selection, playback and exact one-second saved WAV; two export/bounds regressions and nine existing Tauri trimmer tests pass. Both typechecks, Electron lint/build pass. Advanced Tauri pan/handle shortcuts and large-file/native trim interaction still need verification.
Saved voice editor exports portable
.ovsvoicebundles with the same default reference-inclusion choice as Tauri and an explicit preview-only option. Electron routes clone and designed-profile bundles through the shared native Save As IPC path, including authenticated remote-backend requests; the plain-browser fallback retains Chromium download behavior. The native request method is constrained to GET/POST. Real browser/backend round trips pass for both variants, while the focused renderer regression and native controlled-path smoke cover POST export, saved bundle inspection and cancellation-safe UI. Tauri has no standalone inspection UI (its inspect API is shared and covered by the round-trip test), so no additional inspection page is required for UI parity.
Sidebar density correction (2026-09-10)
Nine permanent workspace rows replaced by a keyboard-accessible Base UI Workspaces menu.
Engine status collapsed by default; scrollable details retain all five engine/model settings links. Settings and theme remain on the bottom row.
Removed malformed engine-label separators and reused translated workspace labels across all 21 locales.
Verified Electron typecheck/lint and sidebar-layout-smoke.mjs: 1280x720 and 960x600, nine destinations, five engine links, visible Settings, Stories navigation. Inspected the short-window screenshot.
Navigation feedback is position-stable: workspace, voice/take, engine, Settings, model-family, command and top-bar menu rows use restrained glass/color/shadow transitions without hover translation, rotation or scaling. Semantic chevron expansion remains animated.
Designed voices expose the shared profile editor and persona export pane. The editor includes the missing language choice using the existing searchable language picker, with the selected clone's language synchronized on save. Rename/language persistence and designed bundle export/re-import preserve kind, seed and normalized attributes. Consent recording/revocation, identity lock/unlock, generated preview, synthesis/project/segment usage and guarded project navigation share the same editor for cloned and designed profiles. Persona export now uses the same native Save As contract as generated media rather than an app-scheme blob download.
Shared profile editor now shows synthesis/project/segment usage and project links, with inline confirmation before replacing a dubbing draft and production guards. Backend usage now reads current dubSegments snapshots, supports legacy segments, respects explicitly empty current lists and skips malformed entries. Corrected the shared ProfileUsage response declaration and copied usage translations across all 21 locales. Six offline/empty-cache backend regressions (four failed before) and three UI interaction tests pass; real designed-profile editing still passes. Electron typecheck/lint/build and Tauri typecheck:ci pass. Tauri's broader checkJs typecheck reports widespread errors across application/tests and is not green. Running backend has not been restarted, so the new usage scan is verified in isolated backend tests, not the existing live process. Full native usage navigation remains unverified.
Profile editor includes an expandable Try this voice preview with a ready-to-run localized phrase, independent text/audio, the saved profile's language/seed/style, and paused Vidstack playback after generation. It reuses the existing generation API and a shared interactive synthesis lock without overwriting the clone/design script or Latest take. Closing the pane aborts the request and revokes its audio URL; success refreshes history/usage and can lock the successful take as the profile identity. Six preview/generation tests pass, including lock contention, editor preservation and aborted-pane cleanup. The optional VOICESTUDIO_TEST_PREVIEW=1 real-backend design smoke generated and played a short preview with the installed engine, preserved the design draft, completed export/edit checks and cleaned its history/profile fixtures. Electron typecheck/lint/build pass.
Profile pane now exposes optional consent recording with the existing input/channel controls, localized statement, persisted verification date, inline revoke confirmation and unlock controls for locked profiles. Verification, revocation and unlock each confirm success just as the Tauri profile does. Uses the shared recorder with reference-state tracking disabled for this independent recording purpose. Generated profile previews and take details can lock a successful take as the profile identity. All 21 locale strings are present. Consent/recording regressions pass; a real browser/backend fixture-microphone smoke records, verifies stored statement/status, confirms revocation and cleans the disposable profile. No physical microphone was used. Electron build/lint/typecheck pass. Native unlock/lock interaction remains unverified.
Lock voice identity is now available for a saved-profile preview and Recent takes details, using the exact history ID and returned seed (including zero). The shared backend copies the reference and updates the profile; unlocking re-enables the action. Two lock regressions cover duplicate submission, unknown seed and retry; preview tests remain green. A real installed-engine smoke generated/played a disposable preview, verified locked audio/seed, unlocked through the profile pane and verified both fields cleared, then completed persona/edit checks and fixture cleanup. Electron typecheck/lint/build pass. Native cross-platform lock interaction and full parity audit remain outstanding.
Recent takes details can save generated audio as a reusable clone profile through the existing name/avatar/image form. Take metadata is explicit (complete transcript, language, sanitized style and seed including zero), rather than borrowed from current editor settings; saving leaves the active voice/script untouched. Cancel uses the existing form action, with no additional modal. Real installed-engine generate/save smoke passes and cleans its history/profile fixtures; two API regressions and Electron typecheck/lint/build pass. Unlike Tauri's older 50-character truncation, the transcript stays aligned to the complete saved audio.
Tools now includes an inline A/B comparison workspace for saved profiles and shared Tauri presets. It generates sequentially with a shared phrase and existing production settings, uses the interactive synthesis lock, preserves editor state and provides independent Vidstack players. Leaving the workspace aborts the sequence. Two state/cleanup tests and real installed-engine preset comparison/exclusive playback smoke pass; inspected the rendered view. Electron typecheck/lint/build pass. Long-lived comparison state across workspace navigation and additional native/platform verification remain unaudited.
A/B comparison now retains phrase, voice choices, search filters and completed audio across tool/workspace remounts, matching Tauri's drawer lifetime. An in-flight request is still aborted on departure; replaced audio URLs are revoked by the shared session state. Three component tests pass, including remount preservation; real two-preset generation and playback pass after switching to Probe and back. Electron typecheck/lint/build pass. Persistence is within the renderer session, not across application restarts.
Voice designer now includes the backend personality chips (Narrator, Casual, News Anchor, Storyteller, Corporate and Energetic) alongside its existing presets. Tauri's shared instructToVdStates helper replaces the full attribute set, preventing stale conflicting tokens; manual selection cancels description mapping and preserves text/seed. All 21 personality labels copied. Two component tests and live-catalogue browser checks verify replacement plus script/seed persistence across reload. Electron typecheck/lint/build pass. Demo cards and demo-audio first-run fallback remain separate pending workflows.
Voice Design now offers the seven bundled demo cards, expanded for an empty script and available through Hear demo otherwise. Vidstack previews use local pre-rendered audio; Use this design replaces attributes through the shared normalizer, sets script/language, cancels pending description mapping and collapses the cards. A live bundled Librarian preview and script/attribute handoff pass without generation. Electron typecheck/lint/build pass. The broader setup wizard remains pending.
Cumulative renderer verification: 224 tests across 55 files pass after profile/preview/consent/locking/history save/comparison/personality/demo additions.
Clone's bundled demo profile now offers separate Vidstack demo playback when no TTS engine is available, with a model-settings link and clear pre-rendered label. The common readiness path blocks synthesis, shortcuts and command-palette generation in this state; an engine-ready transition removes the fallback without altering user text. Browser smoke uses mocked engine availability plus real bundled audio, verifies playback/shortcut suppression/transition, and passes. Generation regression tests, Electron typecheck/lint/build pass. New-install setup/bootstrap and actual clean-machine availability remain unverified.
Performance settings now exposes reusable system preflight diagnostics from the existing setup API, only after an explicit System check/Recheck click. Displays every backend finding and suggested fix, retains reports and supports retry. Shared PreflightReport declarations extracted to setup-types.d.ts for both shells. Two interaction tests and live report smoke (eight checks, no mount-time request) pass; Electron lint/typecheck/build and Tauri CI typecheck pass. This is the diagnostics component for setup parity; the first-run wizard's step flow and required-model gating remain pending.
Backend-ready startup checks the existing
/setup/statuscontract once. The first Electron launch enters integrated system-check, models, privacy and dictation steps even when a shared cache already contains the required weights; only the Electron completion marker bypasses later launches. Continue is gated on successful preflight and required-model status, and model status polls only during the install step. Existing installer/model, mirror, privacy and shortcut components are reused, with no automatic downloads or telemetry changes on entry. Packaged Windows custom/default/cancel runtime-location selection is verified without creating a runtime before consent; fully uncached cross-platform installation remains open.Setup's system/model steps now expose expandable Hugging Face token, proxy/mirror and media-tool recovery controls, reusing the same components as Settings. Panels mount only on expansion, so opening setup does not invoke credential/media queries or repairs. Three recovery/gate tests and the browser onboarding smoke pass, including access to the token control without leaving the gate and no automatic model downloads. Electron typecheck/lint/build pass. Actual credential changes, downloads and native executable selection were not exercised in this smoke; their existing settings checks remain the relevant evidence.
Setup privacy presents the Tauri consent copy with equal-weight explicit Yes/No choices and FAQ link, shown only for available, unprompted, opted-out analytics. No default choice or mount-time write; proceeding without an answer leaves the preference untouched. Failed writes retain retry controls and do not mark prompted. Setup hides the duplicate analytics toggle; normal Privacy settings retain it. Electron now reuses Tauri's hardened frontend analytics client: it initializes only after the active backend reports explicit opt-in, disables autocapture/session recording/page views and every remote-code/config product, masks DOM content, filters deliberate properties through the shared allowlist, and stops/resets when consent is off. Both desktop CSPs allow only the fixed EU ingestion origin; external scripts remain blocked.
First-run model selection now shows one cross-family checklist: required models lead, backend-curated optional picks follow, and other supported models fold behind a localized count. Full engine controls remain expandable; normal Settings retains its family catalogue. Shared installer/progress/cancel controls are reused, and no installs run automatically. A regression failed before the change and passes after; three library/gate tests, browser required/recommended/optional visibility and full setup completion, Electron typecheck/lint pass. All 21 locales reuse existing Tauri translations. Actual uncached installs and guided dictation replay remain unverified.
Setup and dictation settings now include the three bundled Tauri replay samples, with the sample catalogue extracted into shared domain data. Vidstack handles preview; explicit Replay submits local WAV audio to the existing /transcribe endpoint and renders its real response. Missing assets hide sample cards, duplicate submissions are blocked, leaving aborts requests, and errors stay inline. Two interaction tests, real bundled playback plus active WhisperX ASR replay, setup smoke, Electron typecheck/lint/build and Tauri CI typecheck pass. All 21 locales reuse Tauri copy. No microphone was used. Native shortcut verification status and actual microphone walkthrough remain pending; replay verifies ASR, not global dictation insertion.
Owner sidebar correction: replaced the workspace popup with persistent full-width compact navigation rows, bounded scrolling on shorter windows, and retained pinned Settings/theme controls. Workspace and engine scroll limits preserve saved-voice space. Browser checks at 1280x720 and 960x600 pass for nine direct navigation links, engine controls and Settings visibility; typecheck/lint/build pass. This supersedes the earlier Workspaces menu design.
Native dictation setup now receives a read-only shortcut-press notification from main; tray actions do not emit it. Settings distinguishes registered/unverified, verified current shortcut and inactive state. Six UI/shortcut tests pass; native end-to-end notification verification remains pending.
Dictation now has Tauri's opt-in dictate-over-playback echo cancellation. Electron persists the preference across windows, tags microphone PCM for the backend NLMS canceller and taps native Vidstack audio/video playback into the shared far-end bus while it is actually playing. Unsupported embedded providers degrade to normal dictation, the default remains off, and the settings copy is translated across all 21 locales. Physical speaker-to-microphone cancellation remains unverified.
Native shortcut verification now passes through the actual Windows helper -> Electron main -> preload -> settings path: unverified status before Ctrl+Alt+F24, verified status after its native press. Existing hold early-release, visible-result restart, toggle stop and disable checks also pass. Test uses a separate temporary Electron profile, simulated microphone input and intercepted transcription WebSocket; no physical microphone or user transcript is involved. Native macOS/Linux and physical microphone checks remain unverified.
Settings now includes a Support view with explicit Ko-fi/PayPal donation choices, optional PayPal amount selection, feature/community/follow/private-security/email/website channels. Tauri and Electron share extracted contact URL constants and existing donation constants; all 21 translations reuse Tauri copy. Native links use the existing external-browser bridge, and no external page opens on mount.
Owner Engine-ready refinement: all five engine icons remain visible with Base UI hover details for engine, full model and current status; each opens its settings family. Expanded engine rows now use a compact aligned single-line layout with no internal scroll area. Browser checks verify hover content, five links and no vertical engine overflow at 1280x720 and 960x600; typecheck/lint pass. This supersedes the earlier engine scroll cap.
Support now includes the shared sponsor roster grouped by configured tier, empty-roster invitation, sponsor inquiry/docs links, and the existing commercial-license copy/benefits with prefilled email inquiry. Licensing destinations are extracted to shared contact constants and reused by Tauri; all 21 locales reuse existing copy. Tests verify the sponsor issue template and mailto subject/body; browser Support navigation, Electron typecheck/lint/build and Tauri CI typecheck pass. No inquiry was sent. Donation goal data and diagnostic-prefilled bug reporting remain pending.
Support donation goal now reuses Tauri's progress loading/normalization, completion logic and extracted money formatter. The existing public snapshot is bundled by Vite and fetched locally with offline fallback; the meter shows amount, remaining goal and supporter count, capped at 100%. Two offline/completed-state tests and browser display check pass; Electron typecheck/lint/build and Tauri CI typecheck pass. No external donation service is queried. Diagnostic bug reporter and broader launchpad remain pending.
Support now has an explicit diagnostic Report bug action. Extracted the existing GitHub report-document formatter for both shells; Electron supplies bounded local system/engine context with an allowlist, and all report-body content passes the shared credential/home-path scrubber. Backend failures still produce a partial report. Native external opening leaves review/submission to the user; no GitHub POST is made. Two redaction/offline interaction tests and 51 existing Tauri bug-report/gate tests pass; Electron typecheck/lint/build and Tauri CI typecheck pass. A native Windows smoke verifies the report's Electron/environment context and exactly one external-open IPC call while intercepting the browser launch, so no report is submitted. Stale-build deflection, reachability and breadcrumb context are completed below.
Report-bug native smoke now verifies app:// Support -> diagnostic report -> external-link IPC with browser launch intercepted and no submission. Context may be partial under its 2.5-second deadline; the smoke does not claim all hardware/engine fields were populated. Browser fallback now reserves its tab during the user gesture, clears its opener before asynchronous diagnostics, and reports blocked popups. Four report interaction tests and Electron typecheck/lint/build pass.
Startup/backend failure screens now expose Report bug, using the cached native stage, managed state, exit code, message and bounded scrubbed log tail even if HTTP context is unavailable. Extracted Tauri's chained-traceback truncation into shared crashReport helpers, preserving original causes. Five report interaction tests (including offline native crash/root-cause redaction), 51 Tauri report/gate tests, Electron typecheck/lint/build and Tauri CI typecheck pass. This captures the current native failure, not a durable previous-session crash marker; stale-build, reachability and breadcrumb parity are completed below.
Explicit bug reports now also fetch the existing last-run-crash endpoint within the shared deadline and include prior shutdown evidence with detection timestamp/version/uptime, separately labelled as a previous unclean shutdown. Shared run-sentinel record adaptation extracted for both shells. Reporting does not acknowledge/delete the record. Six report tests and existing Tauri report tests pass; both typechecks and Electron lint/build pass. This HTTP-backed record is available after backend recovery; native durable death markers while the backend remains down are still not implemented.
Electron supervisor now persists the last three unexpected owned-backend exits in a local version-scoped journal, including timestamp, exit/signal, uptime and bounded log tail. Port collisions/debugger terminations are excluded; intentional shutdowns retain their existing guard. Current-version evidence is included in explicit reports even when HTTP is unavailable. Journal read/write failures do not block startup; in-memory evidence remains. Two filesystem persistence/version/corruption/write-failure tests plus six report tests, typecheck/lint/build pass. Real supervisor-process death/relaunch and UI acknowledgement/history remain unverified.
Real isolated Electron crash/relaunch smoke passes: a disposable owned Python child exits 23 with known stderr; native status and disk journal retain it; relaunch with the same temporary profile restores the exact timestamp/exit/log evidence; intentional app quit while another disposable child is running leaves the journal unchanged. The test uses an OS-assigned unused port and does not interrupt the user's backend. Async readiness assertion was replaced with explicit awaited status polling after it initially raced startup.
Saved native crash evidence is inspectable through inline View crash details in Logs and backend-failure screens: timestamp, exit/signal, version, uptime and scrubbed output, plus explicit reporting. Viewing the evidence now persists an acknowledgement through the trusted Electron bridge, retains the record for reports, clears the repair-attention badge, and leaves a later crash unacknowledged. Real isolated crash smoke expands this view and verifies exit 23 and fixture stderr before relaunch persistence checks.
Source-mode Electron shutdown now closes active renderer-proxy streams before quitting. The native crash/relaunch smoke exposed the prior keep-alive hang, then passed its real window-close path with acknowledged evidence restored across relaunch and no deliberate-quit crash record.
Electron uses a localized recovery view for router errors and an outer React boundary for provider/render failures. It offers retry, expandable scrubbed details, contextual documentation, issue search and explicit diagnostic reporting, with errors retained in the frontend console log. Forced routed-render recovery is browser-verified without a reload. A stale dynamic module receives one controlled renderer reload to refetch its cached rejected import; the same repeated failure then opens the repair dock with scrubbed evidence and no reload loop. Strict Mode regression and browser smokes pass.
Routed-render recovery is browser-verified with an injected failing route, scrubbed expandable details and successful retry without a renderer reload. Error recovery also offers the shared scrubbed GitHub issue-search URL; native/browser external-link behavior is extracted from Support into a reusable component. No issue was searched or submitted during verification.
Error recovery now includes contextual documentation links resolved by Tauri's existing error classifier/catalogue, preserving explicit error-class hints. Two recovery tests and routed browser retry smoke pass; Electron typecheck/lint/build pass. The catalogue is reused directly, so taxonomy and URL maintenance remain shared.
Renderer exception capture now shares one origin-aware noise classifier across Logs, user-facing recovery and Repair Agent automation. A dependency-free bootstrap listener starts before the renderer module graph, then hands startup failures to the same repair boundary once React loads. Electron main journals background promise rejections without terminating the shell, while backend, updater, capture, repair-agent, relocation and window-state publishers use one teardown-safe delivery path so a renderer close/reload cannot become an uncaught
Object has been destroyedfailure. Captured renderer faults no longer add duplicate ChromiumUncaughtentries. Detached route transitions use the same observed task boundary, with a source regression preventing raw fire-and-forget navigation. The live 37-route sweep is clean; focused renderer/main regressions and Electron typecheck pass.Packaged backend startup now contains runtime-location, compatibility and source-staging failures inside the supervisor state machine, preserving the failure in the visible log and recovery UI instead of leaking an unhandled promise rejection from the fire-and-forget app launch.
Persistent workspace shell (2026-09-10)
- Moved the existing voice library, workspace navigation, engine summary and settings footer into the shared app shell. All nine workspaces retain the same sidebar instance and width; saved voice/edit/take actions open their existing cloning views.
- Workspace titlebars share geometry and a collapsed-sidebar toggle. The brand returns to cloning; settings retains its dedicated navigation.
- Verified all nine route transitions preserve sidebar position and DOM identity, active destination, collapsed/reopened navigation and brand return. Compact engine/sidebar smoke passes; 250 component tests pass.
Settings and secondary sidebar polish (2026-09-10)
- SettingsContent fills available width; shared setting rows stack controls in narrow content containers. Settings sections use feature-specific Lucide icons.
- Existing workspace control sidebars share sizing, subtle secondary surface, labeled icon headers and collapse/expand without unmounting their controls. Stories/Audiobook voice rows include profile initials.
- Browser checks cover all eight secondary-sidebar routes at 960/1440 widths and settings at 640/960/1920. Screenshot inspected; 253 component tests, typecheck, lint and production build pass.
Voice conversion workflow (2026-09-10, verified 2026-09-11)
- Tools exposes uploaded/recorded source audio, saved clone target, duration matching, cancellation, transcript and Vidstack playback through existing POST /convert.
- Shares synthesis admission with cloning and ignores obsolete results. Three regression tests pass; the live model-backed conversion is recorded in the next checkpoint.
Conversion recovery verification (2026-09-10)
- Live POST /convert first returned the expected typed 409 for a missing Whisper model. After selecting the already-installed Faster-Whisper Turbo model, the same real path transcribed “Native delivery works.” on CUDA, generated a duration-matched Scarlett clone through OmniVoice, persisted a playable 89,804-byte WAV and returned the complete conversion response. The disposable take and source copy were removed after verification.
- Shared API errors now prefer structured human messages while retaining recovery metadata (fail-before/pass-after regression). Conversion source/voice/duration/result persist in session across navigation; cancel and stale-response tests pass. Typecheck/lint pass.
- Updated longform live verification to select voices within its secondary sidebar, avoiding the newly persistent global voice library.
Multi-chapter and multi-voice live verification (2026-09-10)
- Installed OmniVoice rendered a two-chapter M4B from the Electron audiobook view. Vidstack playback advanced; ffprobe verified both chapter titles, continuous boundaries, and 5.45 seconds of playable output. The live script now asserts export metadata.
- Stories rendered two lines with different existing narrator/character profiles. Verified the actor assignment in the render request and played the returned audio with no renderer errors. No model downloads or microphone capture.
- Longform browser regression passes after scoping selectors to persistent global versus secondary navigation: editing, character assignment, import, voice guard, named project persistence, rendering and independent playback.
Backend restart/status recovery (2026-09-10)
- User reported widespread 502/503 responses during backend restart. Confirmed the replacement backend and all eight affected read endpoints return 200 through port 3902. Previous-session marker records an unclean exit but does not establish its cause.
- Native renderer begins in attaching state until its bridge supplies status. A delayed initial snapshot cannot overwrite a newer pushed state. Translation engine polling now follows backend readiness. Two fail-before/pass-after regressions; 257 tests, typecheck and lint pass.
- The VM134 reportAllChanges stack is not identified in application source; no claim of fixing that separate error.
Responsive dual-sidebar hierarchy (2026-09-10)
- At 1180 px and below, workspaces with local controls reduce the persistent main sidebar to a 48 px icon rail. The rail keeps all nine destinations, active-route state, engine links, settings, theme and the VoiceStudio home action without consuming the editing canvas.
- Wide windows retain the complete saved-voice/library sidebar. Voice cloning has no secondary controls, so its primary sidebar remains full width at narrow sizes.
- Browser regression verifies both responsive states, every workspace destination, secondary collapse/expand, clone return and zero horizontal overflow. Typecheck and lint pass.
Dubbing structural editing (2026-09-10)
- Segment rows now expose insert, delete, cursor-aware split and previous/next merge actions plus 50-step undo/redo and matching Cmd/Ctrl shortcuts.
- Reuses Tauri's shared text-offset attribution model, preserving speaker, profile, direction, gain and target language across merge-to-split round trips. New media, subtitle imports, translations and projects reset stale history.
- Fail-before/pass-after session coverage and the browser dubbing workflow verify every operation, persisted edits and the existing generation/export flow. All 21 Electron locales include the reused translated labels.
- The main editor adds a proportional drag/resize timing lane, keyboard nudging, selection highlighting and overlap warnings. Preparation duration drives its scale; Tauri's shared clamp and speed-recalculation helpers preserve generation behavior.
Model cache location (2026-09-10)
- Settings > Storage can choose or reset the model cache through Electron's native directory picker. Main validates the selected directory is writable and writes the backend's one-shot
models_dircapability; the renderer PUT contains only that token, never a raw host path. - Cancellation performs no write, reset uses an authorized empty path, and the saved restart requirement is shown inline. Native authorization and renderer contract regressions pass. App-data relocation and full cleanup are covered separately below.
Model recommendation and download visibility (2026-09-11)
- Machine-curated models are identified in normal family catalogues as well as first-run setup. Active installs show authoritative aggregate bytes, percentage, transfer rate and ETA from the reconnectable status endpoint, with a determinate accessible progress bar when total size is known. Classified install failures now survive navigation/reconnect and remain attached to their model row with the backend reason, an authoritative retry countdown, contextual Settings recovery and Dismiss.
- A renderer regression covers the recommendation and aggregate telemetry contract. Windows now has an uninterrupted empty-runtime, empty-uv-cache and empty-Hugging-Face-cache packaged proof through a valid paused first-sound WAV and clean shutdown; macOS and Linux remain platform-verification tasks.
Synthesis runtime state (2026-09-11)
- The fixed-width synthesis action now follows the backend model-status state machine: starting the AI runtime, loading weights, warming speech recognition, optimizing the model, generating and receiving audio. Backend model-load percentage is shown beside elapsed time and drives the existing progress track until audio delivery begins.
- Hook and action-bar regressions cover the real sub-stage/progress contract and stable button geometry. All runtime labels are localized in every Electron locale.
Reference transcript integrity (2026-09-11)
- Transcript-free clone references now use the selected installed ASR engine and fall back to the selected installed dictation model before OmniVoice's bundled transcription path. Prompt creation enforces this centrally across clone, streaming, batch, dubbing, audiobook and OpenAI-compatible generation without downloading a model implicitly.
- Both existing empty-transcript clone profiles were transcribed and repaired. Replaying “Native delivery works.” with the original failing seed now transcribes exactly, without the prior leading words. Backend syntax, Electron typecheck and lint pass; the durable backend change takes effect on the next backend restart.
Saved-profile portrait search (2026-09-11)
- Saved-profile editing now offers the same explicit five-result portrait search as profile creation. Uploads and searched thumbnails share one validated image-update path, refresh the profile cache immediately and preserve the existing local initials fallback.
- Keyless portrait search now falls back to Openverse whenever Google's anonymous image-search request fails or returns no trusted thumbnails. The explicit five-choice flow therefore survives blocked and JavaScript-only Google responses without requiring an API key or downloading an untrusted original.
- A component regression covers both sources through the encoded profile image endpoint; no search or upload runs without a user action.
Device-aware model bundle (2026-09-11)
- The Models landing view now uses
/setup/recommendationsto present the backend-curated bundle for the current device before the existing engine-family subpages. - Users can start one model, the required subset or every missing recommendation. Active jobs gate duplicate requests, and the model-cache disk capacity is shown with a warning before the full bundle exceeds it.
- Component and browser regressions cover the low-space state and a scoped one-model install. All visible copy is translated across the Electron locales.
Accurate sidebar runtime health (2026-09-11)
- The sidebar now consumes the actual
/model/statuscontract (status,sub_stage,detail,error) instead of obsoleteloaded/loadingbooleans. - The stable sidebar label is “Local device”; its live dot still reflects backend and selected-TTS health. Engine discovery, missing TTS, model loading and model-load failure have distinct status text and colors; expanded engine rows include individual state dots.
- The device panel reports CPU utilization/model/cores/clock, RAM use, GPU utilization/model and VRAM use, and links directly to remote-backend setup. OmniVoice no longer presents a healthy dot while model status is pending or failed.
- Unit coverage locks the discovery, unavailable, loading, failure and ready transitions.
Backend reconnect query discipline (2026-09-11)
The renderer API now refuses backend calls locally while the native supervisor is not ready. React Query stays offline during startup/crash recovery and resumes mounted queries from the supervisor's ready event.
Backend-unreachable, proxy 502 and shutdown 503 responses are not retried immediately, removing the duplicate request burst that previously flooded profiles, history, engines, model status and settings routes during a restart.
Recent takes now includes both cloning and Voice Design generations with compact All/Clone/Design/Starred filters. Reusing a Design take restores its script, described attributes, seed, profile identity and language in Voice Design instead of incorrectly opening Voice cloning.
Opening either history mode now uses the same integrated take-details pane. Design takes retain playback, transcript/recipe metadata, audio export, save-as-voice and identity locking; Reuse returns to Voice Design with its stored recipe instead of crossing into the clone editor.
Take details now expose the complete reproducible generation record: voice kind and identity, creation and render timing, language, seed, style, every retained production value, duration policy and processing toggles. The compact history list remains scannable while the full recipe stays one click away.
Generated speech never starts without a Play action. Clone, Design, saved-profile, Stories and Dubbing previews remain paused when their artifact arrives, while the main synthesis flow shows localized completion feedback and retains the normal Vidstack Latest take. Cancellation and reported generation progress remain independent from playback.
Voice Design now uses the same fixed composer geometry as cloning: the primary action never shifts as its label changes, model-loading/chunk progress and elapsed time are explicit, cancellation has its own stable control, and long backend errors stay behind a compact expandable summary.
Clone script authoring parity is complete for the current Tauri flow: cursor-anchored expression insertion, clipboard append/replace with undo, and TXT/Markdown/DOC/DOCX/PDF/EPUB import are integrated without changing the expanding editor layout. Clone generation controls retain stable geometry, explicit model/chunk progress and separate cancellation.
Restored the Sharing & Remote Access copy in all 21 Electron locales from the authoritative Tauri locale set after an earlier copy introduced mojibake in non-ASCII text.
Sharing now manages MCP agent voice bindings against the existing loopback CRUD API: users can assign a client ID to a saved or default voice, inspect current bindings and remove them with confirmation. The responsive Base UI form, errors and documentation link use localized copy in all 21 locales; the live backend returned the current binding list successfully.
Settings now includes Tauri's bundled, CDN-free Scalar API reference. It fetches the live local or selected remote OpenAPI document through Electron's authenticated same-origin bridge, rewrites the interactive server target through that bridge, lazy-loads Scalar only on the API page, and retains copy/open/retry affordances. The running app fetched
/openapi.jsonsuccessfully; all surrounding copy is translated across 21 locales.First-run privacy now waits for the analytics availability check and, when consent is available and unprompted, requires one of the same equal-weight Yes/No choices as Tauri before continuing. Unavailable or already-answered analytics remains non-blocking.
Workspace navigation now follows task frequency and product concepts: Stories; Dubbing with Dub/Batch Dub subtabs; Persona with Saved/Gallery subtabs; Transcribe; Design; Audiobook; Projects; Tools. Active subtabs reveal in place, compact mode retains direct access, and every row has fixed geometry with subtle glass, color, ring and icon feedback. Settings navigation and its Back action use the same interaction language. Command search uses the same taxonomy across all 21 locales.
Stories, Batch Dub, Gallery and Transcribe now keep those exact task names in workspace headers, secondary sidebars, the launchpad and project metadata, avoiding the old label change after navigation.
Command search now separates workspaces, current actions, personas and recent takes with icons and platform-correct shortcut cues. Every shared workspace titlebar exposes the same command trigger, while keyboard navigation still operates over actionable rows only and filtering retains the same direct restore/open behavior.
The global
?shortcut and Command menu now open a localized shortcut sheet describing Electron's actual navigation and dubbing-editor commands. It stays inert while typing and uses fixed-geometry hover feedback across all 21 locales.General settings now exposes all 21 interface languages with system-locale detection, durable explicit selection and correct document language/direction for Arabic. Locale bundles load only when selected, reducing the initial renderer JavaScript by about 2.7 MB in the production build.
Updates now renders the shipped local changelog and latest pre-migration database-backup status alongside the live Electron updater. Previously missing update labels, including release-note headings, are populated across all 21 locales in both clients.
Runtime status now distinguishes loaded, selected, inactive, loading and unavailable engines semantically. In particular, the explicit Off LLM no longer appears active and unavailable engines use an error state. Engine summary, icon shortcuts and expanded rows share the sidebar's glass hover/press and reduced-motion-safe icon feedback.
Engine hover details now include the backend's sanitized dependency, routing or model-load reason when an engine is unavailable, while keeping the compact six-icon summary unchanged.
The runtime summary now includes the installed and resident speaker-diarisation model alongside TTS, ASR, LLM, Translation and Dictation. All six families link directly to their model subpage and expose their full model/state details on hover without expanding the sidebar.
Model, translation and sidecar setup failures now share one compact settings recovery surface: a readable summary, dismiss action, exact backend detail and Logs link. Translation and sidecar actions no longer discard the backend cause, and model-family settings navigation uses distinct family icons.
ASR settings now expose the existing OpenAI-compatible transcription backend for OpenAI, Qwen3-ASR, FunASR, SenseVoice and local compatible servers. Base URL, model and encrypted API-key replacement persist through the shared backend; Test saves edited values first, probes without uploading audio and reports latency or an actionable failure. Engine availability refreshes immediately after save, and all new copy is translated across 21 locales.
Transcription history now treats the backend's cleaned transcript as the primary result for lists, copy, export and Voice cloning handoff, matching dictation delivery. The raw transcript remains available in an inline disclosure, search covers both forms, and compact language/time metadata improves history scanning.
Transcribe and Projects now share that same preferred-transcript rule. Sending a result to Voice cloning carries the cleaned text and detected language together, while the original ASR text remains inspectable and copyable in both desktop clients.
Native dictation history now synchronizes through the shared store across Electron's separate capture and main windows. A completed desktop hotkey capture appears immediately in Transcribe and Projects without a route reload; same-window Tauri capture keeps the same subscription contract.
File-transcription failures now keep the primary surface to one actionable sentence with Retry and Dismiss. The backend reason and Logs link remain available through an inline technical disclosure instead of filling the workspace with raw diagnostics.
Persona > Saved now unifies cloned and designed profiles instead of hiding designs. Kind icons make the distinction scannable; selecting a clone restores its reference context in Voice cloning, while selecting a design restores its complete recipe/seed/language in Voice Design. Editing, preview and deletion remain shared, and deleting an active design safely detaches its retained draft.
Larger saved Persona libraries now gain inline search across name, recipe and reference transcript plus compact All/Clone/Design filters; small libraries keep the quiet two-click surface. Profile and recent-take rows use the same glass hover/press and reduced-motion-safe identity feedback as workspace navigation.
Voice Design now keeps its active saved persona visible above the script, matching Voice cloning's stable identity pattern. The identity control opens and focuses the existing saved-design list, so switching voices does not introduce a modal or a second selection mechanic. Its composer also opens the shared production-overrides pane, and generation failures use the same compact summary, technical disclosure and Logs recovery link as Voice cloning.
Editing a designed persona now updates both its validator-safe instruction and complete category recipe. Electron treats a recognized edited instruction as authoritative and uses the stored recipe only to recover empty or legacy-poisoned profiles, so a successful edit changes the next generation instead of silently restoring stale attributes.
Persona preview, ownership verification, revoke and unlock failures now retain the backend reason. Preview errors use the shared compact summary, technical disclosure and Logs link; verification errors remain inline with the failed action instead of replacing the cause with a generic retry label.
Inline transcription model recovery (2026-09-11)
- The Transcriptions workspace now keeps missing-ASR recovery in context: it offers an already-installed dictation model first, otherwise the backend-recommended model, with the remaining catalogue progressively disclosed.
- Downloads start only after an explicit click, persist the chosen dictation model, show real byte progress, support cancellation and automatically unlock upload/record actions when installation completes. An unavailable dictation engine retains the Settings recovery link.
- Component regressions cover installed-model selection and explicit recommended download. The browser transcription workflow now starts from a missing model, installs one inline, then transcribes, exports and clears history.
Designed-voice selection and Dub setup (2026-09-11)
- Selecting a saved designed voice now restores its complete recipe, language and seed, persists that selection, and sends its profile identity during generation so history and usage remain attributed to the chosen voice. Saving a new design selects it immediately; save failures retain the draft and expose the backend reason.
- Dub setup now groups source, language/translation, timing, production and export controls in a wider stable sidebar. Media can be dropped directly, advanced controls remain collapsed, and the empty workspace shows the real three-step workflow without inactive footer actions.
- Interrupted Dub tasks expose Resume, Cancel and Change file in one recovery surface. A backend 404 is treated as already cancelled; an unreachable cancellation can be dismissed or explicitly abandoned locally instead of leaving the workspace permanently disabled.
- Dubbing now retains structured preparation, transcription and generation failures from the backend. The compact error surface shows the sanitized reason and recovery hint, opens the matching troubleshooting page and copies the backend-provided diagnostic;
detail-only transcription events no longer degrade to “Task failed.”
Sharing and remote backends (2026-09-11)
- Settings exposes explicit LAN sharing with confirmation, PIN links, copy/open actions, generated QR codes, configurable share port and Tailscale serve controls. Nothing is exposed until the user enables it.
- Remote backend selection validates the VoiceStudio service before switching. Master credentials are exchanged in main for a scoped memory-only session; renderer requests cannot supply or persist authorization. URL-only preference storage preserves a visible local-backend recovery action.
- Production
app://traffic and Electron development HTTP traffic follow the selected backend through main-owned proxies. Authenticated dictation obtains a path-bound WebSocket ticket from main. Unit coverage verifies credential replacement, streaming proxy behavior and ticket preservation. - Native batch watch folders stream directly to the selected local or HTTPS remote backend and receive only the scoped session header. A real helper smoke verifies confined multipart streaming and authorization replacement; real LAN/Tailscale/remote-machine verification remains pending.
Scoped application reset (2026-09-11)
- Settings > Storage now offers UI, settings, downloaded-assets and full-data reset presets plus exact per-scope choices with real disk sizes. Shared Hugging Face caches are identified before confirmation.
- Content removal requires the localized confirmation word. Main accepts only known scopes from the trusted renderer, refuses remote backends and unsafe filesystem roots, stops the local backend before deletion, and restarts it afterward.
- Generation/dub history, Electron drafts and IndexedDB projects are cleared by their owning layer. Pending and pagehide persistence is suspended so confirmed drafts cannot reappear during reload.
- Thirty-two focused reset, persistence and filesystem tests pass; all reset copy is present in 21 locales.
Verified application-data relocation (2026-09-11)
- Settings > Storage moves voices, projects, outputs and the application database through a native empty-folder picker. The renderer receives a short-lived authorization and never submits an arbitrary host path.
- Main stops the managed backend, copies and verifies the complete tree, atomically switches the shared
OMNIVOICE_DATA_DIRsetting and confirms the restarted backend advertises the destination before deleting the source. Activation failure restores the prior setting and original backend; a separately started or remote backend is refused. - Five filesystem/persistence tests cover successful cleanup, rollback, unsafe destinations and preservation of unrelated durable environment values. Two renderer tests cover the authorized confirmation flow and clear unsafe-folder errors. Relocation copy and progress states are present in all 21 locales.
Native full-data uninstall (2026-09-11)
- Settings > Storage lists every present data, runtime, log and durable-environment location with measured sizes. A shared Hugging Face cache remains unticked until the user explicitly includes it; typed confirmation is required.
- Main rescans instead of trusting renderer paths, rejects remote or separately managed backends, stops the local backend, and gives a minimal exact-path plan to the packaged desktop helper. The helper canonicalizes and revalidates every target, waits for Electron to exit and then removes directories that Chromium would otherwise keep locked.
- Three target-plan tests, two renderer confirmation tests and two native-helper ownership/deletion tests pass. The complete Tauri uninstall copy is reused in all 21 Electron locales. The application binary remains under the operating system's installer/uninstaller.
Installed ASR selection and desktop updates (2026-09-11)
- Installed Faster-Whisper CTranslate2 models can now be selected from the ASR catalogue. The selected repository is persisted, applied to isolated sidecars and reported by engine status; Dubbing offers the installed model inline when transcription finds no usable ASR weights.
- First-install model rows now activate installed dictation choices as directly as installed ASR choices. Selection persists the model, enables dictation, and refreshes readiness, shortcut and Engine Ready state, so onboarding cannot leave a recommended dictation download cached but unusable.
- Model Settings now uses distinct family icons, preserves backend failure details, and marks the active curated TTS submodel as Selected instead of leaving every installed submodel looking unselected.
- Settings now owns stable/preview Electron updates with explicit check, download, progress and restart states. Update manifests are separated by operating system and architecture; install waits for backend/native cleanup and flushes the Dubbing draft first.
- Update installation now observes one shared interruptible-work counter across synthesis, voice conversion, Stories previews and stems, long-form renders, recordings, reference and workspace transcription, native dictation and batch uploads. It also rechecks Dubbing recovery, active model downloads and background batch jobs at click time, so a route change cannot make an unsafe restart look available. Updater status and busy copy is present in all 21 Electron locales.
- The shared release matrix builds and publishes Electron installers, blockmaps and update metadata beside Tauri artifacts, and includes Electron files in stable SHA-256 manifests. Windows NSIS and Linux AppImage/deb packages now build from the current renderer/backend with matching channel metadata; macOS packages and signed update delivery remain unverified.
- Dubbing abort is idempotent when the background task already ended, removing the false 503/retry state caused by the stream and abort request settling in either order.
- The native tray now mirrors Studio essentials: show/focus VoiceStudio, start or stop Dictation with the current shortcut, open Settings and exit through the normal cleanup lifecycle. It uses the existing high-resolution brand tray assets, switches to the recording indicator while listening, follows the active app language and disables Dictation while a transcript is still being finalized.
First-use cloning demo (2026-09-11)
- Selecting the bundled demo voice with an empty script now supplies a localized starter prompt once and explains the guided state in the composer. User edits immediately dismiss the guidance and existing text is never replaced.
- When no TTS engine is usable, the normal fixed-width synthesis action becomes a Vidstack-backed Hear demo control instead of presenting a disabled dead end. Playback still uses the application-wide single-player policy, the model-install explanation stays in context, and the real synthesis control returns automatically when an engine becomes ready.
- The global synthesis shortcut is explicitly inert during this fallback. The prompt is present in all 21 Electron locales; Electron typecheck passes.
Existing-install analytics choice (2026-09-11)
- Configured installs that predate onboarding now receive the same one-time, equal-weight analytics choice as fresh installs. It appears as a compact glass banner only when the backend reports analytics available, off and never answered; onboarding installs keep the choice inside their privacy step without duplication.
- Yes, No and dismiss all persist an explicit choice through the existing backend contract. Dismiss keeps analytics off and stops future prompting; an unreachable backend leaves analytics off and allows the question on a later launch.
Shared long-running failure recovery (2026-09-11)
- Dubbing, Stories and Audiobooks now preserve the backend's structured public failure payload instead of reducing it to a generic error. The shared compact surface presents the sanitized reason and remediation, opens the mapped troubleshooting page and copies the sanitized diagnostic.
- Dubbing also accepts
detail-only transcription failures, so an actionable ASR message no longer degrades to “Task failed.” - Long-form cover upload, chapter planning/preview, project storage, line auditions, stem rendering, file import and export now retain their actual failure message behind the same dismissible surface instead of showing a bare “Error.” The Stories speed control also renders its multiplier correctly rather than displaying replacement question marks.
- Batch Dubbing now preserves the first per-file enqueue cause while retaining only failed files for retry. Cancel, delete, export and queue-loading failures show the actual action-specific error, with retry or dismiss in the shared compact surface.
- Persona Gallery archetype use, Community use/design/submission, imports and catalogue loading now retain the actual backend or storage cause. Failed catalogue requests expose Retry while failed user actions can be dismissed without clearing the current search or selection.
- Tools utility, voice-conversion and voice-comparison failures now use the shared compact recovery surface. Conversion retains missing-ASR model recovery, comparison preserves the backend cause for the failed voice, and utility responses no longer reduce a returned error to a generic label.
- Projects now identifies the failing library or save/open action while preserving loaded content and offering an in-place query retry. Dubbing export and native watch-folder failures retain their filesystem/backend cause and can be dismissed without resetting the Dub or Batch Dubbing workspace.
Pronunciation dictionary and workspace navigation (2026-09-11)
- Settings now exposes the shared pronunciation dictionary with global and language scopes, respelling, IPA and CMU entries, enable/delete controls, a live substitution preview and JSON add-or-replace backup restore. The complete surface is localized in all 21 Electron locales.
- Workspace navigation uses glass hover/press feedback and reduced-motion-safe icon motion. High-frequency destinations lead the list; Dubbing groups Dub and Batch Dub, while Persona groups Saved and Gallery without changing their existing routes or workspace state.
Native permission visibility (2026-09-11)
- Settings now reads microphone authorization from Electron on Windows and macOS, reports Linux honestly as unknown, and rechecks when the app regains focus. A denied microphone opens the matching OS privacy pane directly.
- macOS also reports the Accessibility grant needed for dictation delivery and opens its exact System Settings pane. Permission IPC is limited to the trusted main frame; browser-only rendering explains that native state is unavailable. The complete surface is localized in all 21 locales.
Local usage insights (2026-09-11)
- Settings now presents the backend's existing local history aggregates: generated takes, produced audio, voices, active days, dubs, favorites and mode/language distributions. Empty, loading and exact failure/retry states stay inside the page.
- The page states its privacy boundary in context: it creates no new history, storage or telemetry and reads only the local backend. A remote session offers an explicit return-to-local action instead of making a false local-only claim. Usage copy is translated in all 21 Electron locales.
Voice preview cache consent (2026-09-11)
- Models now exposes the same consent-first pre-rendered voice-preview cache as Tauri. No outbound preview fetch starts while disabled; enabling it downloads the featured set and Check now bypasses the normal refresh throttle.
- Cache readiness, localized relative freshness and exact rejected-update details stay in the settings surface. All nine preview strings are present in every Electron locale.
Model and saved-profile completion audit (2026-09-11)
- Model installation status now performs one initial read and polls only while a nonterminal install job exists. Live idle observation recorded zero status calls over ten seconds while the native health monitor retained its intended two-second cadence.
- Model Settings keeps the voice-preview consent control under one visible heading; its accessible row label remains available to assistive technology without repeating the same title inside the single-row card.
- Reference details respond to the actual pane width instead of the window width. Narrow voice-sample panes stack Transcript and Style, and the ASR transcript uses a content-sized multiline editor so long recovered speech remains readable and correctable without widening the inspector.
- Direct comparison with Tauri's VoiceProfile flow confirms the shared Electron editor covers editing, reference playback, preview generation, portrait upload/search, usage, consent, identity locking, persona export and deletion. Clone and Design open that editor in a dedicated wider, independently resizable pane; its content becomes a two-column layout only when the pane has room.
- Blank reference transcripts are resolved during profile creation through the selected installed local ASR/dictation engine. The router now imports the async runtime it invokes; previously the missing import forced every automatic transcription attempt into the best-effort fallback, allowing the first synthesis to inherit stray reference words.
- A disposable live clone saved without reference text recovered
native delivery works.through the selected Faster-Whisper Turbo model, returned the populated profile, and removed both its database row and owned audio on deletion. - Transcription history now selects the newest visible entry automatically, moves selection to the first matching result when filtering and shows entry count plus clip duration in the secondary sidebar. Existing history no longer opens into a contradictory empty-detail state.
- Choosing a new uploaded or recorded reference now clears the prior saved profile's transcript, language and delivery direction together with its identity. A new voice therefore starts in Auto without silently inheriting metadata from the previous persona.
- Appearance keeps the branded palette's stable internal compatibility key while presenting its polished product name,
VoiceStudio Classic, instead of exposing the raw legacy identifier. - Model-family pages now reserve their row geometry with shared settings skeletons while engine/provider state resolves. Installed ASR models keep their selection action disabled until the active checkpoint is known, so the currently selected Turbo model never briefly offers a contradictory Select action during startup.
- The static model catalogue now refreshes through its install, repair, delete and recommendation mutations instead of polling every 30 seconds while Settings is open; live install-job and residency polling remain scoped to their changing runtime state.
- Voice-profile name copy now distinguishes labels from placeholders in every Electron locale: persistent labels and accessible names are plain nouns, while only empty text fields carry the trailing ellipsis.
- Transcription readiness now polls only while the selected Fast or Accurate model is unavailable, including the bundled replay used by setup and Dictation settings. A background install still unlocks each workflow automatically, while ready Parakeet and Faster-Whisper sessions stop issuing permanent idle readiness requests.
- Engine Ready now refreshes resident models once per second only during active work and immediately at each activity boundary. A live Fast transcription showed Dictation loading, retained Parakeet TDT v3 on CPU, then settled directly to loaded without the prior fifteen-second false-idle window.
- Engine Ready now treats the shared lifecycle inventory as the authority for residency instead of cached engine evidence. Engine/model selection, removal and manual unload invalidate both inventory and engine status, preventing a previous model from leaving a newly selected or unloaded engine falsely green.
- Engine Ready follows foreground model requests across Clone/Design generation, saved-profile previews, A/B comparisons, voice conversion, Audiobook chapter previews, Dub segment previews and Dub QC. Their existing cancellation and failure paths release activity exactly once, so the fixed engine bars reflect work outside the main composers too. A live Scarlett generation showed TTS amber while working, then green with OmniVoice resident on
cuda:0; its disposable history item was removed. - Designed profiles whose deterministic identity sample was deferred now retain a preview action in Saved, Design and the profile editor. Playing it reaches the backend's existing lazy-render path, reports TTS activity, refreshes the profile after success and then behaves as ordinary playback instead of leaving the recovery path unreachable. A disposable pending profile was exercised through the live Saved sidebar: its preview request returned audio, persisted the generated identity WAV and produced no renderer errors; the row and file were removed afterward.
- Profile deletion now clears both cloned-voice selection and the active Voice Design persona through the shared mutation, including deletion from the editor pane. Saved Voices no longer duplicates that mutation's success or failure notification.
- Clone readiness now includes the selected TTS engine and its explicit model installation state. Missing or incomplete weights keep Synthesize disabled with a direct TTS setup action, closing the remaining predictable click-then-model-error path; the old Electron-preview wording was removed from recovery copy.
- Voice Design now consumes the same TTS/model readiness without inheriting Clone's reference or script blockers. Generate remains stable and disabled while engine state resolves or weights are unavailable, with the same direct TTS setup action instead of a predictable backend failure.
- Profile test generation, A/B voice comparison and deferred designed-persona sample playback now share that TTS readiness source. Existing saved audio remains playable without a model, while synthesis-backed controls stay disabled and point to model recovery. The first-run sound request waits across the short readiness-query race instead of being silently discarded.
- The shared model catalogue now waits for the managed backend's ready stage before requesting state. Reusing it for synthesis readiness therefore preserves the setup shell's no-premature-backend-request contract.
- TTS readiness follows the selected engine rather than any installed alternative. Composers, profile previews, recovery notices and readiness polling now stay blocked when the active engine is unavailable, matching the backend's explicit no-fallback generation contract.
Diagnostics and support parity (2026-09-11)
- Settings includes the Tauri self-check report with actionable deep links, copyable scrubbed diagnostics, and a locally generated diagnostic bundle that opens in the file manager.
- Sidebar and settings navigation hover states preserve their geometry. Feedback is limited to color, glass, ring, shadow and icon treatment, so repeated pointer movement does not shake surrounding content.
- Saved-persona rows leave their full width to the voice name and age at rest. Preview, edit and delete appear as a stable glass action cluster on pointer hover or keyboard focus, preserving direct access without permanently truncating ordinary profile names.
- The primary workspace sidebar now owns a bounded full-height flex layout. Its voice list yields vertical space so the persistent workspace, Engine Ready and Settings controls remain visible at high DPI or UI scaling.
Dubbing casting parity (2026-09-11)
- Dubbing now exposes one project-level Casting Board for every detected speaker. Saved voices and design presets can be selected with the keyboard or dragged onto a speaker; assignments update every matching segment and retained merged-speaker attribution through the existing undoable, persisted edit path.
- The pristine Dubbing workspace now includes Tauri's bundled local source/dub comparison with four target languages, synchronized Vidstack playback, a direct upload action and durable dismissal. It fetches only packaged demo assets from the active backend.
- That bundled comparison now eagerly loads only its local MP4 pair, so both cards show real duration and a ready frame on arrival instead of two black
0:00players. The shared VideoPlayer keeps click-to-load as its default for remote and ordinary media.
Dubbing stage guidance (2026-09-11)
- Completed transcription and translation stages now surface a compact, dismissible checkpoint with the segment count and one clear next action. Completed generation reports timing overflow segments for review without hiding the editable timeline.
- The pristine Dubbing canvas now explains the three-stage workflow without repeating the source sidebar's file action. Additional target languages use removable chips and no longer claim that zero languages are selected beside the already-selected primary target.
- Source-language and speaker-count controls now live with the source upload they affect instead of under the target-language settings. The same pre-upload disclosure reports whether the speaker-diarisation model is installed and links directly to its model page; missing diarisation remains an explicit, visible quality fallback rather than blocking single-speaker work or appearing only after transcription.
- Electron now consumes the shared system-notification contract in the Engine Ready area. Compact actionable notices surface missing credentials/media tools, low storage, GPU compatibility and prior crashes without adding a modal; actions route to the matching settings view, crash notices acknowledge through the shared endpoints, and dismissible standing notices persist locally while errors remain visible until fixed.
- Stories character stems now follow the selected export format: MP3 uses the shared
/stories/encodemedia path, while M4B mode retains editable WAV stems. An unavailable encoder falls back to WAV with the same localized notice as Tauri instead of failing the completed render.
Audiobook playback and live render feedback (2026-09-11)
- Stories now exposes only its character cast editor; the separate
[voice:NAME]cast mapping remains scoped to Audiobook scripts, removing two adjacent Cast sections with overlapping purpose. - Audiobook renders now show per-chapter pending, rendering, cached, complete and failed states with elapsed time and ETA. Stop settles into a clear stopped state, while validation warns about empty chapters, unknown voice tags and missing cast assignments before work starts.
- Completed books retain the exact rendered script and chapter map. Vidstack playback follows the current word, chapter and timestamp; clicking a word or chapter seeks directly to it.
API reference and engine visibility (2026-09-11)
- Settings > VoiceStudio API keeps Scalar on the existing settings route, shares the renderer's single React runtime and disables its network-backed agent/developer surface. The embedded reference now opens without hook errors, route loss or blocked outbound requests.
- A compact 1280×720 sweep now covers Appearance, Models, TTS, Compute Device, Privacy, Storage, System Preflight and VoiceStudio API with no route errors or page overflow. System preflight uses the shared responsive settings row instead of overlapping its explanation and action; Diagnostics also tolerates a temporarily unavailable preload bridge and its live self-check reports 10/10 healthy.
- Settings > Sharing now reaches the backend MCP binding contract through its complete
/api/mcpnamespace. The binding list loads without a false 404; create and delete use the same corrected route family. - The sidebar engine strip now reports live loading, loaded, idle, inactive and failed states with fixed colored bars under every engine icon. Dubbing transcription, translation and generation drive ASR, Translation and TTS independently; fast/local transcription and live capture drive Dictation without falsely lighting the selected ASR. Resident models resolve from the lifecycle inventory, and hover details expose their actual execution device.
- Background Batch Dubbing jobs now drive those same engine bars from their authoritative pipeline stage: transcription lights ASR, translation lights Translation and generation lights TTS. Upload, extraction and muxing no longer falsely report TTS work, and residency polling stays responsive while a batch continues off-page.
- Engine Ready now polls only live model/runtime state. Static engine, translation, dictation and installation catalogues refresh from their owning selection, credential and install mutations, cutting the idle sidebar refresh burst from six backend requests to two every fifteen seconds while preserving immediate status changes.
- OmniVoice startup residency now feeds the engine catalogue even before a synthesis route creates its adapter.
/enginesand/model/loadedagree on the loaded CUDA device, while engine evidence also exposes the model's actual float16 precision. - Engine health resolution, model-family metadata and the shared catalogue query now live outside React component modules, so development updates to runtime and model views preserve the current workspace instead of invalidating and remounting the application shell.
- Model Settings now matches resident dictation weights by their runtime model ID as well as repository ID. Parakeet therefore exposes its real in-memory state, execution device and unload action instead of looking merely installed while it occupies memory.
- Engine selection pages now keep usable engines first and fold unavailable engines into one disclosed group with their recovery reason. Platform-incompatible curated choices stay hidden until their parent engine is available, and unavailable rows no longer present inert Select actions.
- Dictation no longer repeats its section label around the enable switch, Diarisation owns a correctly named/iconed model surface, and LLM provider setup omits the empty file-model catalogue. All six family pages mount without renderer errors.
- LLM Providers now uses one Base UI provider selector instead of seventeen competing buttons. Active, configured and local state remain visible; provider guidance, signup/help, credentials and model controls use the shared settings-row hierarchy, with local providers correctly omitting the API-key field.
- Translation Settings presents ready engines first, folds credential-dependent engines into one unavailable group with their recovery reason, removes inert selection actions, and avoids a duplicate LLM Providers route at the bottom of the page.
- Permissions and Dictation settings retain their accessible row labels while rendering only one visible page heading, removing duplicate labels from the compact settings hierarchy.
- Launchpad saved-persona actions preserve voice kind: cloned voices open Voice cloning, while designed voices restore their complete recipe, seed and language in Design.
- Primary workspaces and Settings now use TanStack Router's supported lazy route components while the shell and launchpad remain immediate. Intent preloading warms a workspace on pointer or keyboard intent before navigation. The production startup renderer fell from 4.47 MB to 1.60 MB before compression; all ten workspaces and seven model/settings routes still mount without console errors, recovery screens or horizontal overflow at 1280×720.
Transcription quality routing (2026-09-11)
- The Transcribe workspace now exposes a persistent Fast/Accurate choice without changing live Dictation. Fast routes uploaded or recorded clips through the selected low-latency dictation model; Accurate preflights and runs the selected ASR with detailed timestamps. Each path has independent readiness and opens the matching model family when unavailable. Engine Ready attributes the work to Dictation or ASR accordingly.
- Fast file transcription now shares live Dictation's silent-model recovery. If a selected sherpa model loads but emits no tokens, VoiceStudio retries only through an already-installed local ASR, adopts the recovered transcript and demotes the broken runtime only after the fallback proves the clip contained words. Uploaded transcription and automatic saved-profile text therefore cannot silently accept an empty decoder result or trigger a model download. The active Parakeet model transcribed a saved Russian reference in 0.41 seconds through the live endpoint without needing recovery.
- Its empty state has one message in the content canvas while the history rail retains only its zero count. Record is a neutral capture action until recording begins, so it no longer looks selected beside a primary Start dictation recommendation; active recording receives the destructive stop treatment without moving the toolbar.
- A live Accurate upload through the Electron renderer exercised the selected Faster-Whisper Large-v3 Turbo model, displayed amber ASR activity while working, returned a timed transcript in 6.3 seconds, released the request-scoped ASR runtime afterward and produced no renderer errors.
- Gated diarisation installs now expose both required Hugging Face access actions, local token settings and immediate retry on the model row; the raw Hub response is retained only under Details. A gated 403 can no longer leave its README-only cache marked installed.
speaker-diarization-community-1is the preferred future local pipeline, but it requires pyannote.audio 4 while the Python 3.11 WhisperX build requires pyannote.audio 3.x; move diarisation to an isolated runtime before replacing the compatible 3.1 pipeline. - Fast/Accurate now names the exact selected Dictation or ASR model beside the quality control. The workspace and Engine Ready share the same cached dictation selection, so changing models cannot leave the two surfaces describing different work or add a duplicate catalogue request.
- Missing-model recovery follows the selected quality path: Fast offers Dictation setup, while Accurate offers ASR setup. An unavailable low-latency model no longer interrupts a ready Accurate workflow with unrelated installation UI.
First-install setup completion (2026-09-11)
- A fresh setup owns the full window, so saved voices, workspace navigation and command search no longer mount behind it or issue premature backend requests. Existing installs still enter their workspace without an extra screen.
- Command search remains available from Settings after onboarding; only the setup-owned window suppresses it.
- System preflight now runs on entry and gates Continue from its real result. OS permission state is visible during setup, media tools remain invisible when ready and become directly recoverable when acquisition fails, and UI scale can be adjusted before model installation.
- Every setup step now owns the same full remaining window width, so moving between System check, Install models, Privacy and Dictation keeps the content column centered without shifting or narrowing it.
- Setup-in-progress state suppresses the legacy analytics banner while the Privacy step owns that choice. A browser audit confirmed one inline choice and no overlay during setup, while configured pre-onboarding installs still receive the one-time banner.
- Completing setup now persists independently from model readiness. Deleting or moving a required model later keeps returning users in the studio with normal model-recovery controls instead of reopening first-run onboarding. A first Electron launch always completes system, model, privacy and dictation setup even when it discovers weights in a shared cache; cached weights alone no longer masquerade as completed onboarding.
- Resetting all VoiceStudio settings and content clears the Electron onboarding markers before reload. The documented “fresh install” reset therefore returns to system/model/privacy/dictation setup, while narrower UI, settings, asset or history resets keep the completed first-run state.
- AudioSeal watermarking now explicitly converts each bounded chunk to the model's required 16 kHz rate. Embedding resamples only the generated watermark residual back onto the untouched source-rate speech, and detection uses the same native-rate contract; 24/44.1/48 kHz outputs no longer rely on AudioSeal 0.2's ignored legacy
sample_rateargument. - Every successful Electron native media save now enters the shared export history. Clone/design takes, Dubbing outputs, Batch Dubbing outputs, Stories and Audiobooks therefore appear immediately in Launchpad and Projects with their actual chosen filename, destination and media type; history bookkeeping remains non-fatal after the file is safely written.
- Settings > Performance now exposes the live loaded-model inventory, individual unload actions, allocator-cache flush and one guarded Unload all action. The shared backend action now fulfills its label by releasing TTS, alternate engines, dictation, translation, diarisation and sidecars instead of unloading only OmniVoice; active Electron work disables destructive memory actions.
- The packaged runtime stages the frozen Python project, checks 9 GiB of free space and write access, downloads a private pinned uv only after the install action, supports a custom location and retry/cancel, preserves compatible Tauri data/runtime state, and separates model downloads into explicit user actions. Windows completed a fully uncached packaged run through a valid paused first sound; macOS and Linux remain release-gate work.
- Completing onboarding now opens Voice Design and creates the localized first sound through the shared generation controller. Cold-model stages, cancellation, errors, output history and Vidstack playback therefore use the same mechanics as normal synthesis; only clean-machine cross-platform proof remains a release gate.
- Installed Faster-Whisper choices can now be activated directly from their first-run model row, so an explicit optional download does not leave the new model hidden behind advanced engine controls. Setup navigation uses the same family icons as the workspace while preserving explicit downloads and TTS-only entry.
- Runtime setup reports resolved dependency count, announced/downloaded bytes, rate, remaining data, ETA and the largest active package. Windows snapshot creation preserves the canonical Hugging Face blob through a hard link or copy fallback, preventing model loading from repeating the completed multi-gigabyte transfer.
- A packaged runtime left with its install marker now returns to the same setup surface as an explicit Resume operation. The existing uv project and cache remain intact, clean reinstall remains separately confirmed, and startup still never downloads without the user's action.
- Shared missing-TTS notices now offer one Fix action: it selects an installed compatible engine or starts the recommended required-model install, shows live progress without accepting duplicate clicks, activates it when the job completes, and hands only failed or unexplained recovery to the opted-in local repair agent with an app-level action request.
- Dubbing and Voice Conversion retain their direct installed-ASR and Models recovery paths, while their no-model state and Dictation catalogue/install failures now also offer one explicit agent Fix action carrying the precise setup task and current retry context.
- Voice Design, voice comparison and saved-profile previews use the same in-place TTS recovery. Checkout-free agent operations discover supported endpoints from the scoped bridge, choose ordinary hardware-compatible setup values and verify readiness; licenses, credentials, privacy consent, deletion and remote-backend connections still require the user.
- Missing reference ASR, accurate upload transcription and live-dictation models expose the same app-operation repair action while retaining the selected audio, transcript edits and retry state.
- The current Windows packaged profile now opens past Setup directly into the launchpad with its generated “Welcome aboard…” first-sound take visible and no renderer error; live setup status reports every required model ready and the existing analytics choice remains explicitly prompted.
- Repair agents can operate on the backend currently attached to the app through a per-run loopback capability bridge. Electron main retains the real local or remote backend target and authentication, Diagnose is enforced read-only, auth routes are blocked, and teardown removes the bridge and capability file; focused bridge tests pass.
Voice Design conflict recovery (2026-09-11)
- Manual accent and dialect choices now explain which conflicting category was cleared instead of silently changing another control. Reset to description immediately reapplies the current mapped recipe, and resets all attributes to Auto when the description is empty, matching the established Tauri behavior.
- Clone and Design generation failures now share the compact dismissible recovery surface. Long backend traces collapse to their actionable first sentence while the complete sanitized diagnostic remains available through Copy diagnostic instead of expanding the composer or duplicating the error in a toast.
Explicit translation model lifecycle (2026-09-11)
- NLLB-200 now appears in Settings > Models > Translation with the shared install, progress, repair and delete lifecycle. The engine remains unavailable until its complete pinned model is cached, and translation loads local files only, preventing an unannounced 2.4 GB download.
- A retained NLLB runtime now appears in the shared loaded-model inventory with its actual device, can be unloaded from Model Settings, and drives Translation's ready/idle indicator in Engine Ready. The normal post-job release also clears stale device state.
- Alternate TTS models held by the local generation cache now join the same resident inventory and unload contract. Graceful backend shutdown releases shared TTS, alternate engines, sidecars, dictation, diarisation and translation before allocator teardown instead of relying on process exit.
- Argos now reports exact source-to-target pack readiness without contacting its package index. Electron and Tauri disable translation for missing pairs and expose the download as an explicit install action; the translation route returns a typed 409 and leaves the package state unchanged instead of silently downloading on first use. A live missing
en → decheck and installedpt → entranslation both pass without changing installed packs. - Translation readiness now means the dependency/model and required provider configuration are both present. Settings, Dubbing, Batch Dubbing, Engine Ready and the backend routes share that contract, route setup to Credentials or LLM Providers, and prevent click-then-fail jobs.
- Batch Dubbing snapshots the selected translation provider and uses the same Argos/NLLB/online/LLM dispatch as interactive Dubbing. It no longer calls Google regardless of the user's choice or silently labels untranslated source speech as a completed target-language track.
- License-gated Supertonic-3 and PocketTTS engines now expose the shared local acceptance flow in Electron before installation or selection. Both license sets and PocketTTS's Hugging Face access conditions open through the trusted native browser bridge; acceptance refreshes engine availability without starting a model download.
- Engine family rows now expose their effective device, stable install guidance, path setup snippet, trusted documentation and an explicit health/re-check action. Probes remain click-only, refresh live engine residency after completion and keep the default model page compact by folding unavailable engines into one disclosure.
- Failed Batch jobs retain their app-owned upload for one-click recovery in both desktop clients. Missing Argos packs identify the exact source/target pairs; Install downloads only those packs, then retries the same job. A live
pt → derun reached the typed setup state, retry remained a 409 without installation, no pack was downloaded, and the job/input were deleted after inspection. - A live MP4 run now verifies the selected Faster-Whisper Large-v3 Turbo ASR, NLLB translation, saved-profile OmniVoice synthesis, muxed output and owned-file deletion end to end. The run also fixed saved-profile rows crossing into synthesis as
sqlite3.Rowobjects and silently producing an empty target segment. - The separate interactive Dubbing contract is also live-verified: MP4 preparation, exact Turbo transcription, NLLB Spanish translation, saved-profile OmniVoice generation, persisted translated segment/track, MP4 export with H.264 video plus 24 kHz AAC audio, history deletion and complete job-directory cleanup all completed successfully.
Backend supervision and performance profiles (2026-09-11–12)
Electron no longer kills a still-running managed backend after three health-probe timeouts. A controlled Windows run suspended the idle Python worker through three misses: Electron reported temporary unavailability, retained the same PID, restored Ready after resume, and left the crash journal unchanged. Actual child exits retain crash reporting.
Once the backend has reached Ready, that temporary-unavailable surface now overlays the mounted workspace instead of replacing it. React Query remains offline until recovery, while route components, scroll positions, selections, and unsaved editor state retain their component identity. Initial startup and real child-crash recovery keep the existing full gate. Electron typecheck passes.
The desktop parent-pipe watchdog retires its own backend sentinel before its deliberate containment exit. Controlled Windows restart and parent-pipe checks leave deliberate exits clean; this does not establish that every historical unclean shutdown was a false positive.
Diarisation candidates remain Community-1 through an isolated pyannote 4 runtime and Sortformer v2.1 through a compatible native runtime. Upstream audio.cpp currently documents v2.1 as local conversion only pending redistribution approval; do not advertise a downloadable v2.1 Q8 package. Existing pyannote 3.1 compatibility remains required.
Controlled Windows parent-pipe verification confirms a live child creates its sentinel, exits with code 0 after pipe closure, and removes its sentinel. This used an isolated temporary data directory and did not restart the user's backend. Electron production build also passes; the managed health-timeout recovery path is verified above.
Performance preferences persist the global tier and family overrides as one atomic object. A concurrent six-family write check retained every override; a global reset removed them in one write.
Startup profile reconciliation preserves a ready explicit network translator, but replaces an unavailable one with the tier-appropriate installed local provider. The live Max profile recovered from unavailable Google to ready NLLB without downloading weights; ten focused offline profile checks pass.
Model installation/selection and translation-provider changes now invalidate Performance state immediately, so its effective model and applicability rows cannot trail Engine Ready by the five-second poll interval. Ten focused renderer regressions, Electron typecheck and targeted lint pass.
Engine Ready keeps all six family/model pairs visible at the normal 231 px row width, truncating both columns without horizontal overflow. Its sidebar smoke now guards the current six-engine and Stories navigation contract instead of the obsolete five-engine/Stories Editor labels.
Settings navigation contracts to 208 px below 760 px while retaining search, every route and full hover labels. The Models recommendation smoke guards the reclaimed canvas and zero horizontal overflow at 640 px.
Profiles owns the Saved library at every window width and keeps the main library as a compact rail, so Saved voices appears once instead of as identical sidebar and content lists. The normal rail toggle still restores the full library on demand.
The compact rail's top control now uses Lucide's explicit sidebar-open icon, making the existing expansion action recognizable instead of presenting an app-logo button with hidden behavior.
Performance preset resolution lives in
services/performance_profiles.py, shared by API clients. OmniVoice generation, Dubbing, Batch, Voice Conversion and long-form rendering consume Fast/Balanced/Quality/Max defaults only after an explicit preset choice; request-level Production overrides retain priority. Untouched preferences preserve historical defaults, and other TTS engines receive no guessed sampling settings.Engine Ready includes the four-level stepped speed/quality control; Performance settings expose the global preference and all six family rows. Controls disable while foreground or Batch work is active, and unsupported selected runtimes remain visibly unavailable.
Faster-Whisper maps preset decoding effort to beam/best-of widths 1/3/5/8 while preserving the selected model, VAD and timestamps. Max may select the largest already-installed compatible ASR model but never starts a download.
Preset mutations now share a serialization scope and busy state across sidebar/settings instances. Required engine reads and lazy workflow imports finish before preference persistence, preventing a failed prerequisite from reporting a partially applied selection. Background Batch work and unavailable batch status also disable preset actions through the shared active-job query. Electron typecheck passes.
NLLB translation honors Fast/Balanced/Quality/Max through 1/3/5/8 decoding beams. Options are captured once per job and reused on its MPS-to-CPU retry, so segments retain consistent effort. This preserves the explicitly selected provider and all source/target languages; no LLM provider is enabled and no model is downloaded by the preset. A live installed-model run verified Fast at one beam and Max at eight against the same segment.
Preset controls consume the backend's implemented-family list; LLM remains visibly unavailable instead of accepting no-op changes. Diarisation becomes applicable only with both local runtimes installed. Failed preset saves retain the requested tier for a real retry, while failed batch-status reads retry that prerequisite independently.
Diarisation installation now validates configuration-only snapshots against their catalogue-required files instead of rejecting them for containing no large weights. Installed-state and download validation share the same helper, including rejection of empty/missing required files. Python compilation passes; gated download and dependent-weight installation remain unverified.
The existing pyannote 3.1 install now includes pinned segmentation and speaker-embedding dependencies through the same explicit install, cancellation and retry path. Preflight includes their bytes, and the catalogue marks a config-only cache incomplete when dependencies are absent. The shared token resolver is now used by standard downloads and their preflight, matching the accelerated path. Python compilation and YAML loading pass; no model weights were downloaded for verification.
Pyannote runtime configuration now resolves the installed pipeline and both checkpoints locally, rewrites their references in a temporary config, and publishes the resident pipeline only after device placement succeeds. Missing cache files no longer initiate a nested Hub download during a job. Python compilation passes; full diarisation remains pending gated installation.
Dubbing now distinguishes missing diarisation checkpoints from runtime load errors and points missing-cache recovery to model installation/repair. Installed bundles no longer require a retained HF token. Generic Hub errors are no longer automatically classified as license failures. Python compilation passes; actual gated model execution remains unverified.
Crash-isolated Faster-Whisper receives the same bounded ASR decoding options per job as the in-process engine. The child validates the options and retains historical defaults when absent; a live installed-model run produced the same saved-profile transcript in-process and through the isolated child. Sherpa transducer dictation now implements the same four tiers through greedy and modified-beam search.
audio.cpp catalogue readiness now requires the exact Breeze Q8_0 package, rather than accepting unrelated weights from the shared multi-model repository. Required checkpoint files use the shared per-format size floor; install errors name missing required files. Python compilation passes. This hardens package identification but does not implement additional native model families.
Performance controls distinguish implemented families from the selected engines that actually consume presets. Unsupported selected runtimes disable their family control; the global control disables when none apply. The profile query refreshes availability after engine changes.
The preset API now rejects unsupported selected engines and changes while persisted task jobs or Batch jobs are queued/running, covering another client bypassing the renderer controls. Python compilation passes. This is an admission check, not an atomic lock spanning every generation path; in-flight decoding still relies on per-job option snapshots.
Controlled native restart completed through Electron's normal window-close path after the backend reported idle and both job lists were empty. The app relaunched, OmniVoice preloaded successfully, and the live preset endpoint now returns applicable_families=[tts] with the existing quality preference preserved. The new startup log contains no unclean-shutdown/sentinel warning. This verifies one orderly restart, not all historical crash causes or health-timeout recovery.
The opt-in native Sortformer v1 adapter feeds audio.cpp sample-indexed turns into the shared annotation pipeline. It uses installed local files, the existing native device resolver and process containment; exact-speaker-count requests fail into the established warning/fallback path rather than being silently ignored. The app-managed v0.7.4 runtime and Q8 model processed a real 60-second, two-person Dubbing source on the Windows RTX 4090 Vulkan device in 3.84 seconds, returning 16 bounded turns across four detected speaker labels.
scripts/smoke_sortformer.pymakes the local-media check repeatable without downloads or app-data writes.Native Sortformer registers its owned process with the shared dubbing process registry, polls cancellation while waiting, and unregisters on every exit. Dubbing discards completed diarisation output when the job was aborted rather than persisting it after cancellation. The repeatable smoke can interrupt real native inference through
--cancel-afterand verifies that no process remains registered.Native diarisation output now carries audio.cpp Sortformer attribution into downstream speaker-clone processing instead of being labelled pyannote. The adapter rejects non-GGUF inputs before native launch and retains a bounded native failure-log tail in backend diagnostics before temporary cleanup. The real Windows Vulkan inference above verifies the native output path; other desktop runtimes remain release gates.
Speed/quality uses a real Base UI stepped slider with four circular stops, a draggable icon thumb and a filled range. Dragging previews locally and saves on commit, preserving the existing busy/error guards. Engine Ready also has a Lucide CPU icon beside its live status dot.
The four-level slider was inspected in the running renderer with the current API; active and unavailable family states render correctly. Native Sortformer missing-file recovery names its CLI/GGUF setup instead of directing users to install pyannote.
Install cleanup now initializes its heartbeat stop event before fallible setup and stops it unconditionally in finally. Early token/import/revision failures retain their original diagnostic instead of raising an unbound-local error inside the error handler. Python compilation passes.
Sortformer preprocessing no longer reads entire recordings into NumPy or makes full-size resampling copies in the backend. Compatible mono 16 kHz PCM WAV input is reused directly; other supported inputs are normalized by a contained, cancellable ffmpeg process. Annotation bounds use file frame metadata. The 60-second live-media smoke verified normalization, inference and bounded annotation output; maximum-length memory measurement remains a release check.
Diarisation has a read-only selected-runtime status endpoint used by Engine Ready. Native Sortformer reports its configured model and missing binary/model guidance; pyannote readiness checks the complete dependency bundle. Existing loaded pyannote memory cannot be misattributed to selected Sortformer. The live Windows endpoint now reports the selected Sortformer runtime and model ready, matching the successful native inference smoke.
Engine Ready now shows an actual Working state while native diarisation owns a preprocessing/inference process, with an amber activity bar and translations in all 21 Electron locales. Process activity is cleared on success, cancellation and failure; idle native adapters are not reported as resident loaded models. TypeScript/Python compilation plus real Windows Vulkan inference and cancellation pass; other desktop runtimes remain verification gates.
Sidebar sizing constrains the main grid column and footer to the resized width and truncates engine names without pushing status/actions out of view. Speed/quality uses one keyboard-accessible four-level range with a value-dependent background.
Diarisation recovery: runtime load failures now report DIARIZATION_LOAD_FAILED instead of a licence failure and link to general recovery. Documentation matches explicit installation and offline inference; native Sortformer remains opt-in and now has real Windows Vulkan inference evidence.
Speed/quality reference correction: the separate radio cells became a continuous rounded four-stop range with a filled accent track and icon thumb. It supports pointer dragging and keyboard adjustment while retaining persistence and busy guards.
Shared error documentation maps now include missing diarisation files and runtime load failures in both backend and Tauri; existing taxonomy expectations updated without running suites.
Read-only renderer inspection: speed/quality renders with aligned thumb in Settings. At 196 CSS px content width and 150% element zoom, clientWidth and scrollWidth both measured 196 (no horizontal overflow); current Quality endpoint was visually inspected. No preset was changed. Full sidebar inspection in a fresh browser context was blocked by onboarding, so this does not establish the entire native sidebar at zoom.
Persona > Saved now routes to /personas, a responsive searchable saved-voice library using shared preview/edit/delete/selection behavior. Choosing a voice opens its clone/design workflow; simply opening Saved no longer opens the composer. Narrow sidebar tabs preserve the active label, engine rows prioritize family names with model details on hover, and library search stays pinned within the independently scrolling list. Electron typecheck and native Saved-library editor/export interaction pass.
Voice cloning keeps its main-sidebar toggle beside the page title, with the correct open/closed icon, persistent workspace state, and a localized accessible label/tooltip.
Saved-library follow-through: profile editing stays in the shared resizable editor alongside the library; selecting a voice clears that editor before opening its generation workflow. Command search > Persona > Saved uses /personas too. Electron typecheck passed; no tests run.
Saved-library browser inspection against current local backend: /personas rendered eight real profiles; opening an edit control kept #/personas and raised no page errors. No profile was saved, deleted or generated. The disposable browser context alone used a completed-onboarding marker; native app onboarding/preferences were unchanged.
Runtime feedback: active TTS/ASR, translation and dictation operations use Working rather than implying model loading throughout. OmniVoice's explicit loading response retains Loading; model-status polling accelerates while foreground work runs. Reuses existing translated Working state. Electron typecheck passed; no generation or tests run.
Reference transcription requests selected ASR (reference mode), not the separate dictation engine; shared backend installed-model preflight remains authoritative, word alignment/refinement stay off, and manual-edit race guards are preserved. The selected installed Faster-Whisper Turbo path subsequently completed live reference transcription without downloading a model.
Speed/quality feedback now increases accent saturation, fill highlight and restrained glow across Fast/Balanced/Quality, with a matching tier icon. Theme primary tokens drive the colors; disabled controls suppress glow, reduced-motion preferences disable transitions, and geometry stays fixed. Typecheck and narrow-width Quality visual inspection passed.
Compact sidebar libraries: saved voices and recent takes use single-line 36px rows, with metadata in hover titles and actions revealed on hover or keyboard focus. Takes preview through the shared audio button; full waveform/details remain in the take pane. Virtual history estimates now match compact rows. Full Profiles library retains its detailed cards. Typecheck passed.
Automatic reference ASR requires confirmed installed files at endpoint preflight and loader entry, including fallbacks and preflight failures. Strict checks refresh cached installed state. Successful selected Faster-Whisper reference inference is recorded below.
Reference transcript recovery exposes a localized missing-ASR explanation and direct ASR-model settings link alongside Retry. Uses existing translations and native link semantics. Electron typecheck passed. Notifications now live beside Settings; floating theme control was removed at owner request.
Engine hover tips use theme popover tokens and a shared structured body for family, live status, selected model, execution device and recovery detail. Browser focus/hover inspection confirmed the themed popup; initial TTS/ASR/LLM lookup now shows Loading rather than a premature unavailable/inactive state. Typecheck passed.
Translation, dictation and diarisation tips now distinguish initial lookup from inactive state and failed status reads. Dictation keeps its selected model visible even when disabled, rather than replacing model identity with Inactive. Electron typecheck passed.
Engine tips anchor to the shared icon row in the expanded sidebar and the shared engine group in the compact rail. Browser inspection measured identical x=12 and width=231 across TTS/ASR/translation/dictation in the expanded sidebar; content remains theme-aware and width-contained. Typecheck passed.
Reference ASR distinguishes a failed installation check from confirmed missing files; generic conflicts no longer become misleading missing-model guidance. Verification failures remain retryable and do not initiate downloads. Python compilation and Electron typecheck passed.
Controlled app restart after confirming no active jobs: normal window close shut down the previous processes/listeners; dev app relaunched successfully. Fresh health returned ok, TTS status ready, reference mode present in running OpenAPI, and diarisation endpoint reports pyannote not installed/not loaded. No unclean/crashed/sentinel line appeared in the fresh startup log; the subsequent installed Faster-Whisper reference request completed successfully.
Live reference-mode request using a 1.97-second existing local take returned typed 409 asr_model_missing for selected WhisperX/Systran large-v3. The installed-only guard stopped before inference; no model downloaded, profile changed or take created. This verifies missing-model recovery, not successful transcription. Selected ASR preferences were preserved.
Successful live reference transcription: selected the already-installed deepdml Faster-Whisper large-v3 Turbo via the normal selection API after checking job queues idle; CUDA routing confirmed. Reference mode returned HTTP 200 from faster-whisper in 2.65s on the existing 1.97s take, with text 'High Native Delivery Works.' (intended 'hi Native delivery works.'). This verifies execution/selection, not perfect recognition accuracy. No model download or new voice/take was created. Installed Turbo remains selected.
The complete uploaded-reference save path now passes through the live Clone UI. A 16.6-second MP3 transcribed through
mode=referencein 2.37s; upload, name, portrait and Save then returned a selected profile with the 291-character transcript and normalized image, without renderer errors. Direct profile, audio and JPEG reads also returned 200; both disposable profiles were deleted with their owned media. No model or runtime download started.Gated-model guidance is now a compact disclosure beneath its model controls. Access links, token setup, recheck/retry actions and raw errors appear only after expansion. Browser inspection confirmed the pyannote card starts collapsed with no page errors; Electron typecheck passed.
Model catalogues now use responsive auto-fill cards with installed, selected, resident-memory and download state kept on each card. Active cards receive a restrained accent surface, narrow views collapse to one column, optional/incompatible groups retain responsive grids, and Hugging Face mirror presets are full-width selectable menu rows with endpoint/status detail. Electron typecheck passed.
The sidebar performance control is now the rail alone, with a smaller high-contrast thumb and speedometer, small sparkle, full sparkle or crown for Fast, Balanced, Quality and Max. Saturation and glow rise by tier; a shared anchored engine-style tooltip explains the current choice. Max applies 64-step TTS and stronger local decoding, then activates the largest already-installed compatible ASR, dictation and NLLB models without downloading anything. The live backend advertises all four tiers; Electron typecheck and targeted Python compilation pass.
Standard Ctrl/Cmd
+,-and0shortcuts now step or reset the same persisted UI-scale setting exposed under Appearance, including numpad and non-US keyboard equivalents. Browser zoom is suppressed so CSS scale and Settings never disagree. Native Windows capture verified Ctrl+− reduced the complete shell and Ctrl+0 restored 100%; Electron typecheck passes.Diarisation now exposes persisted pyannote/audio.cpp Sortformer engine selection in Settings. The reviewed audio.cpp bundle includes Breeze-TTS-2 and Sortformer Q8, the native adapter resolves only installed cache files, Engine Ready follows the selected runtime, and switching to native releases resident pyannote memory. The live status endpoint returns both options; TypeScript/Python compilation and catalogue parsing pass. Managed native inference is verified below.
Diarisation install, repair, reinstall, delete, recommendation, and background-completion paths now invalidate both its Settings selector and Engine Ready status. A completed explicit install therefore exposes Select immediately without a reload or idle poll. Electron typecheck and lint pass; no model download was started.
A missing selected diarisation runtime now carries its actionable recovery reason through the live status endpoint, Engine Ready tooltip, and Settings engine row while retaining the exact model identifier. Fresh native restart returned
Install the pyannote model bundle; health stayed OK and startup recorded no unclean warning or error. Python compilation, Electron typecheck, and lint pass.Diarisation readiness now separates installed Sortformer weights from the required audio.cpp executable. Settings and Engine Ready retain the exact model identity and report the missing runtime action instead of contradicting the installed Model Library card or reducing the cause to a generic unavailable state. Four focused readiness regressions, the route contract, Python compilation, and Electron typecheck pass.
The pinned audio.cpp runtime is updated from v0.7.2 to v0.7.4 with release-published SHA-256 values on Windows, Linux and macOS. Settings exposes an explicit managed install with streamed progress, size/checksum verification, traversal-safe extraction and a native device probe; the installed runtime survives app updates and requires no environment variable. Its v1 adapter opts into graph growth, clamps at most one decoder frame of final padding to the media boundary, and rejects clips beyond the supported 120-second path with direct pyannote guidance. The managed Windows Vulkan runtime completed a real 35-second inference with nine turns/four detected speakers; the 12-second clip and process cancellation proofs above used the same release. Sortformer v2.1 streaming is documented but remains local-conversion-only because upstream has no redistributable GGUF package.
Fresh development startup resolves the app-managed audio.cpp v0.7.4 runtime without environment variables. The live installer endpoint reports managed Windows x64 readiness, the diarisation endpoint reports both model and runtime installed, and native Sortformer is now the persisted active engine after confirming every generation queue idle. The running Settings row renders
AvailableandSelectedwith no page errors; backend health remains OK on the RTX 4090 and the latest startup sequence contains no unclean-run warning.The main sidebar's top control now collapses into the existing 48 px micro rail instead of removing navigation. Workspace headers no longer duplicate the logo/settings controls in that state, and collapsed secondary sidebars retain their workspace icon as a clear reopen target across Stories, Dubbing, Profiles, Transcribe, Design, Audiobook, Projects and Tools. Electron typecheck passes.
Runtime device labels now use local NVIDIA, AMD, Intel and Apple marks for CUDA, ROCm/HIP, XPU/oneAPI/OpenVINO and MPS/Metal, with the generic CPU icon retained as fallback. These marks appear in engine hover details and engine-selection device badges; no icon CDN is used at runtime.
The 48 px rail reserves its bottom controls at short window heights: workspace links scroll independently, while a single live engine-status control, Settings and notifications remain reachable. On Home, the collapsed-state header puts the sidebar-expand button directly beside VoiceStudio instead of repeating the brand icon. Electron typecheck passes.
Sherpa transducer dictation now consumes Fast/Balanced/Quality/Max as greedy 1-path, greedy 4-path, modified-beam 4-path and modified-beam 8-path decoding. The selected language model stays fixed except for the existing explicit Max installed-model policy, and a tier change rebuilds the warm recognizer on the next session. The live selected Parakeet TDT v3 Quality path transcribed the complete 16.98 s Scarlett reference in 0.66 s; the endpoint reported dictation as implemented and applicable. Twenty-three focused Sherpa tests, Python compilation, Electron typecheck and lint pass.
Diarisation presets now map Fast/Balanced to installed native Sortformer and Quality/Max to installed pyannote. The family control is enabled only when both are installed, so a single-runtime machine cannot save a cosmetic no-op choice; environment pins remain authoritative and switching to native releases resident pyannote memory. Three focused policy regressions join the 23 Sherpa tests; no model was downloaded or selected during verification.
Main-sidebar resizing now measures the first real layout ancestor instead of BackendGate's zero-width
display: contentswrapper. Pointer and keyboard resizing work after explicitly expanding the library beside a secondary workspace; the edge hit target stays inside the clipped sidebar. The responsive nine-workspace smoke also covers the 48 px rail, wide layout and settings width.Notifications remain visible beside Settings even with zero notices, with a localized all-clear popover instead of removing the bell. Compact and expanded sidebar states share the same control.
The Models landing page uses responsive auto-filled family cards and wrapping recommendation cards, preserving install and active-state controls at narrow widths. Support is a highlighted permanent destination directly above Back; its donation, sponsor, commercial-license and contact surfaces use distinct icons and responsive content. Browser inspection at 1440 px, Electron typecheck and lint pass.
electron/tests/route-layout-smoke.mjsnow makes the live renderer audit repeatable: all 38 workspace, Settings redirect and model-family routes open at 1440 px and 640 px without a recovery screen, page/console/transport failure, backend 5xx response or document-level horizontal overflow. Route-cancelled React Query reads are distinguished from real transport failures; no workflow mutation is performed.electron/tests/accessibility-smoke.mjsaudits every Electron workspace and Settings route at 1440 px and 640 px for computed accessibility-tree names, unnamed controls, missing image alternatives, app-owned duplicate IDs and broken ARIA references, nested interactive elements and missing main landmarks.bun run smoke:serverprovides its repeatable same-origin renderer/backend harness. Engine details remain a valid controlled region while collapsed, OpenAI-compatible ASR fields and narrow icon actions keep accessible names, and Scalar's embedded landmark is named without nesting a second main region.The staged Clone chooser now keeps its Back action beside “Choose a voice” in both states: initial setup returns to the launchpad, while changing a ready voice returns to the existing script. A live renderer interaction verified the initial route transition without changing profiles or reference media.
The current Windows unpacked package was rebuilt after the accessibility, Clone and dependency updates. Its executable exposes the VoiceStudio waveform artwork as its native associated icon; the real
app://voicestudiorenderer opened the existing saved-profile Clone flow, entered and left the chooser through Back without a recovery view or console/page error, and retained a healthy CUDA backend. It was then relaunched normally with no debugging port left open.Focused model-management, runtime-status and reference-transcription regressions are current again: model settings and Engine Ready tolerate incomplete/version-skewed engine payloads instead of crashing or claiming readiness, while automatic profile transcription tests exercise the authoritative single
/transcribepreflight path, typed missing-model recovery, edit-race protection and unmount cancellation. Eighteen focused tests, Electron typecheck and lint pass.The current Electron production bundle builds successfully with Vite 8.3: main, preload and all renderer routes compile, including the responsive Models/Support surfaces and the hardened runtime/profile paths.
Shared backend contract verification now matches shipped behavior: install recovery exposes bounded retry timing, translation selection requires a configured ready provider, and blocked/JS-only Google portrait results fall back to five normalized safe Openverse thumbnails. Ninety-four focused model-management, profile-image, reference-transcription, performance-profile and engine-shape tests pass with
HF_HUB_OFFLINE=1and an empty Hub cache.The Windows unpacked Electron package builds with the branded
VoiceStudio.exe, bundled native desktop bridge and application archive. The post-build artifact contract passes for executable entry syntax, packaged Python resources and the package-version source.A fresh isolated packaged profile reaches the explicit first-run runtime-install action without creating a runtime, contacting Hugging Face or attaching to the development backend. The packaged native bridge also answers its protocol handshake; the multi-gigabyte uncached install and first generated sound remain release-gate work.
Real installed NLLB execution now verifies the translation performance range end to end: Fast resolved one decoding beam and Max resolved eight against the same English-to-German segment. Both returned a complete translation, no download started, and the original global Max preference with no family overrides was restored. Warm-cache timing is intentionally not used as a quality or speed claim.
Real saved-profile transcription now matches exactly between in-process Faster-Whisper and its crash-isolated sidecar using the selected installed Large-v3 Turbo model at Max effort. The run exposed and fixed two integration faults: the sidecar lacked the shared eager-loader contract, and its child rejected Max's eight-beam options. Immediate retry after a child EOF now also retires the stale Windows process handle. Nine isolated-ASR regressions pass offline; the live comparison restored Faster-Whisper and unloaded its sidecar. The unpacked Windows artifact was rebuilt with these fixes, then passed the artifact contract and isolated no-install first-run smoke.
The packaged first-run surface now keeps VoiceStudio branding visible, shows checking, uv acquisition, dependency installation and verification as real progress, and streams the latest installer activity. Failed setup offers Retry and a confirmed Clean & Retry; cleanup is restricted to the default runtime or a custom runtime Electron owns, protecting reused Tauri environments. A rebuilt isolated Windows package completed the explicit install and reached a live same-origin backend.
Clean runtime recovery serializes competing setup actions, publishes filesystem failures with access guidance and logs, and refuses to reinstall after shutdown invalidates the attempt. Seven focused supervisor tests, Electron typecheck and targeted lint pass. The refreshed Windows artifact includes the recovery changes and passes isolated first-run setup.
Interrupted Electron runtime installs are explicitly marked and excluded from both readiness and Tauri compatibility checks until installation and import verification finish. A targeted interrupted-install/retry regression passes alongside the supervisor checks (16 tests); typecheck and targeted lint pass.
Current Windows artifact includes both cleanup serialization and incomplete-install markers; artifact contract and isolated branded first-run smoke pass. Packaging reached executable signing but the redirected shell command returned exit 1 without an error diagnostic, so clean packaging-command success remains unproven. Development app restored with the updated main/preload; live model, profile-image, translation and settings requests return 200.
macOS Electron packaging now carries Tauri's microphone purpose description and explicit audio-input entitlement for the app and inherited helpers. The packaging contract checks both audio access and JIT entitlement; native macOS recording remains unverified.
Direct Node child-process capture of electron-builder completed with code 0 and no signal, resolving the earlier redirected-shell exit uncertainty. The refreshed Windows package passes isolated first-run smoke.
Setup now reserves the operation before asynchronous compatibility checks, preventing duplicate install requests and stale preflight after shutdown. Each attempt resets elapsed time and activity. Eight focused supervisor regressions, typecheck and lint pass; this follow-up is source-verified.
Installer progress labels now wrap without truncation on narrow windows; short-window overflow keeps branding and setup controls reachable. The onboarding status-loading/retry view also retains VoiceStudio branding. Typecheck and targeted lint pass; native visual verification is pending.
Browser inspection at 480x420 renders four installer phases with no horizontal overflow or page errors. Inspection exposed corrupted verification translations from shell encoding; all 21 were rewritten with Unicode-safe input. Installer text now explicitly uses the semantic foreground token for visible branding. Browser status was simulated; no runtime installation was started.
Installer recovery is reachable at 480x420: Clean & Retry opens its confirmation, without page errors or document overflow (simulated backend status). A wider locale audit found 317 corrupted strings containing repeated question marks; 60 were restored from Tauri translations with identical English source text. The 257 strings remaining at that checkpoint were repaired in the completed catalogue pass above.
Dubbing sidebar now grows to 24rem, transcript content uses more available width, unchanged source text is not repeated above its editor, and cast chips wrap. Typecheck and targeted lint pass. Locale recovery additionally repaired 107 navigation, appearance, profile and runtime labels; the 150 strings remaining at that checkpoint were repaired later.
Responsive Dubbing inspection at 1440/960/640px found no horizontal overflow. At 640px, controls now stack above the transcript: editor width increased from 112px to 384px while preserving the wide desktop sidebar. Typecheck passes.
Repaired 68 corrupted model-access/retry messages across nine locales, preserving the seconds interpolation. All locale JSON files parse; the 82 strings remaining at that checkpoint were repaired next.
Repaired those remaining 82 profile, disk-space, performance and player strings.
node electron/tests/locale-encoding.mjspasses across all 21 locales and guards repeated question marks, Unicode replacement characters and provider-error artifacts. The broader catalogue completeness audit is now closed above.A stricter audit found 274 later labels where shell encoding had replaced one or two non-ASCII characters with literal question marks. The affected navigation, theme, profile, model, performance and repair-agent copy is restored across 17 locales; the build-time guard now rejects this subtler corruption while allowing an idiomatic sentence-ending question mark.
Product-name review repaired another 92 localized strings that translated or omitted VoiceStudio, Hugging Face and other engine/tool names. The same build-time guard now preserves those names exactly whenever the English source uses them; all 21 rendered languages and the representative 960/640 px long-string routes remain overflow-free. The installed Windows Sortformer v1 path also completed its full supported 119.5-second input in 6.5 seconds, returning 40 bounded turns across four speakers; inputs beyond 120 seconds continue to require pyannote or locally converted Sortformer v2.1 streaming.
Interpolation audit found five data-relocation summaries displaying a fixed 12 MB instead of the actual size. Arabic, French, Hindi, Russian and Ukrainian now preserve
{{size}}; the locale guard checks interpolation against English across all 21 files, allowing localized zero plurals to omit the count. The guard passes.Locale encoding/interpolation validation now runs automatically before Electron builds and unit tests, and within the packaging contract used by release builds. Production build and packaging contract pass on Windows; the build reports an existing ineffective dynamic import for
dub-session.ts.Live localization rendering now switches through all 21 bundled languages and verifies direction, translated headings and document width. German, French, Russian, Ukrainian, Hindi, Arabic and Traditional Chinese also cover General, Performance, Models, Clone and Dubbing at 960 px and 640 px without renderer errors or horizontal overflow. Saved-voice metadata and reference-ASR links use restored middle-dot separators; Transcribe clear confirmation keeps its count without the corrupted question-mark separator.
Tauri's persisted Dubbing review mode is now available under General. Pause for review keeps the staged ASR/translation checkpoints; Run straight through suppresses those nudges without removing the explicit workflow actions. The setting renders and survives reload, its reviewed Tauri strings are present in all 21 Electron locales, and the expanded General page passes the long-string layout smoke.
Settings > Performance now shares the Local device telemetry contract instead of reducing hardware to RAM, allocated VRAM and a GPU boolean. Responsive cards show CPU model, physical cores, threads, live frequency/utilization; GPU model, utilization and used/total VRAM; RAM; platform, architecture and Python. Live Windows rendering reported Ryzen 9 7950X (16C/32T, 4.5 GHz), RTX 4090 and 23.99 GB VRAM with no overflow or renderer errors. CPU-only torch builds also discover AMD/NVIDIA/Intel adapters from bounded native OS probes, and Intel XPU memory uses torch's native runtime when available; focused Windows/Linux fallback checks pass.
Reset interface now clears every renderer preference, including language, all pane widths, transcription mode, dismissed notices and onboarding hints. It preserves drafts, transcripts and local app settings; All settings additionally resets Electron's local dictation/AEC preference. Three focused reset regressions and Electron typecheck pass.
Partial or version-skewed engine payloads are now normalized through one guarded family lookup. Model Settings, Engine Ready and performance controls keep rendering when any TTS, ASR or LLM family is temporarily absent instead of dereferencing it; seven focused regressions, Electron typecheck and lint pass.
Reopening Settings now restores the last valid page, including model-family pages, matching the established Tauri navigation behavior. Unknown or obsolete stored paths fall back to Appearance, and Reset interface clears the route memory. Nine focused checks, Electron typecheck and lint pass.
Settings > VoiceStudio API now gives Scalar the complete remaining content width and height instead of a padded, fixed-height card. Its responsive toolbar exposes live backend state and endpoint, refresh, copy-spec and raw-spec actions; live rendering at 1440 px and 960 px filled the pane without document overflow or renderer errors. Voice cloning also exposes the shared main-sidebar expand/collapse control beside its title; a live browser check closed and reopened the sidebar while retaining the workspace. Electron typecheck and lint pass.
Structured Electron pipeline failures now preserve the backend's exact contextual
docs_urlinstead of discarding it and relying only on message classification. Only HTTP(S) destinations are accepted; malformed or unsafe values fall back to VoiceStudio troubleshooting. Error-boundary and pipeline-doc routing checks pass (4 tests), with Electron typecheck and lint clean.Diagnostic reporting now consumes Electron's existing signed-updater verdict before gathering or opening anything. A known newer build offers View update or File anyway using the established localized Tauri copy; filing records the stale-build context, while updater failure remains non-blocking. Nine report-flow checks, Electron typecheck and lint pass.
Diagnostic reports now include connection mode, the last backend response observed before report collection, structured transport metadata and a privacy-safe in-memory ring of fixed route/action labels. Clone/design synthesis, Dubbing upload/URL preparation/transcription/translation/generation and Fast/Accurate transcription record start plus terminal outcomes without retaining user text, paths or URLs. HTTP responses, including error statuses, update contact time while transport failures do not. Twenty-five focused report, client and breadcrumb checks pass; Electron typecheck, lint and production build are clean.
Persona bundles now use Electron's native Save As IPC with a constrained POST request instead of Chromium's download manager. A disposable clone completed native save, backend inspection and cleanup; three focused renderer checks, typecheck and formatting pass. The refreshed Windows unpacked package includes this change, passes the artifact contract, reaches branded first-run setup without creating a runtime before explicit consent, and mounts its renderer/preload/Models route against the live backend.
Native Save As now accepts only the trusted main frame and confines authenticated GET/POST downloads to the active backend origin. Renderer API paths, packaged app URLs and same-origin absolute URLs remain supported; cross-origin and non-HTTP targets fail before any request. Five focused export/boundary checks and Electron typecheck pass.
Renderer-generated transcription TXT, pronunciation JSON and Stories stem files now use a separate trusted native Save As IPC path; browser builds retain their ordinary download fallback. Native byte-for-byte save, renderer routing, typecheck, production build and formatting pass. The refreshed Windows package includes both native-save paths and passes isolated first-run plus same-origin live-backend smokes.
Windows native picker coverage now includes packaged first-run custom destination selection, restoration to the default and cancellation without starting installation; FFmpeg executable selection; and model-cache directory selection/reset. Scoped authorization files are verified and removed without changing backend preferences.
Dubbing responsiveness and preview transport (2026-09-12)
- Long Dubbing projects virtualize the transcript editor and draw the timeline from 2,048 backend-cached waveform peaks instead of decoding the full separated-vocals WAV in Chromium. A 155.2 MB real vocals track produced its peaks and onsets in 0.83 seconds; later mounts read only the small JSON cache.
- Dubbed-video previews are written atomically with MP4 fast-start metadata, served with byte ranges and immutable revisioned cache URLs, warmed after generation, and swapped through one retained Vidstack player. Native MIME hints select the provider without a probe, while original and generated extensionless routes answer metadata-only
HEADrequests without reading media or starting a mux. On the real 390.6 MB project, the first mux completed in 17.3 seconds and a repeat range request started in 0.004 seconds. - Local NLLB translation widens batches from live accelerator headroom while bounding the batch multiplied by beam count. A 24 GB CUDA card now processes up to 24 Fast rows or 12 five-beam Quality rows per forward-pass group instead of the former fixed 8/3 widths.
- The Dubbing secondary sidebar grows to 32rem where space permits, replaces the loaded upload form with one source row, and keeps source, target and casting as the visible 1/2/3 path. Automatic casting starts folded. The retained Vidstack player preserves time and play state across Original/Dub changes; its glass controls auto-hide during playback and expose seek, volume, rate and fullscreen controls. Thirty-eight targeted backend regressions, 23 Dubbing renderer tests, locale checks and Electron typecheck pass.
- Dubbing transcript rows now use compact, content-sized translation editors with source/status metadata and speaker/actions grouped into two scan lines. Subtle dividers, subdued metadata and a focused accent edge keep long transcripts scannable without turning every segment into a card. Multi-target projects expose translated-count language tabs above the transcript for one-click editing, while retaining every selected target for batch translation and generation. Legacy provider error pages saved as dialogue recover to the original source and a retryable error badge.
- Agent fit closes the translation-to-speech timing loop in both desktop shells: contextual Autofit creates the first draft, Dubbing measures the real natural duration before any slot padding or trim, and a configured LLM rewrites only untouched lines outside the 0.90–1.04 timing window. Each candidate is rendered again before another correction, with two rewrite passes maximum; manual edits are never overwritten and extremely short lines keep honest silence. The Tauri command bar exposes the same single action and multi-language loop; 49 focused workflow checks and a live renderer smoke pass.
- Translate with Agent remains visible in Dubbing's persistent action footer once a transcript exists instead of requiring Translation quality, Agent fit, or a scroll position. Agent and the active Google/Argos/NLLB/API engine are explicit translator choices; every translate entry point follows that choice, while Fast/Autofit/Cinematic remain engine-quality controls. Packaged Codex, Claude Code and OpenCode each returned complete ordered translations for real time-budgeted segments; Pi remains gated on an installed host. Shared secondary panes apply the same heading contrast, icon emphasis, active navigation, responsive width and inspector hierarchy across every workspace.
- The linked LLM setup now puts provider and skill readiness before legacy engine diagnostics, defaults an unconfigured local-first install to its local provider choice, and makes the primary save action activate both the provider and the matching app engine. Provider changes refresh Engine Ready immediately; a focused interaction regression and a live responsive renderer check cover the one-action handoff.
- Packaged Electron no longer crashes when a launcher or exiting backend closes a diagnostic, child-output, or drain pipe while logs are still arriving. Expected Windows asyncio transport resets are also excluded from repair diagnostics; real stream failures remain visible.
- Electron's lifetime backend supervision now polls the compact canonical
/healthresponse instead of repeatedly serializing/system/info; hardware and settings details remain demand-loaded by their views. The focused supervisor contract and renderer Unicode guard pass. - Electron's explicit runtime installer now matches Tauri's CUDA transcription prerequisites: it detects a live CUDA build, installs and validates the pinned cuDNN 8 compatibility libraries, and clears CTranslate2's obsolete Linux executable-stack request before accepting the runtime. The readiness stamp invalidates older incomplete managed environments without letting them bypass through legacy Tauri reuse, and partial compatibility installs remain recoverable instead of being marked ready. An empty-cache Ubuntu 26.04 WSL install produced a 7.9 GiB environment, reached healthy on the RTX 4090 with Faster-Whisper available and no Hugging Face download, then shut down cleanly; 13 focused installer tests, typecheck and lint pass. A clean AppImage profile also completed explicit runtime installation from the dependency cache and connected its packaged renderer to the live backend. A fully uncached package-to-first-sound run on Linux/macOS remains a release gate.
- Runtime setup now parses uv's streamed package and byte progress into the native status contract. The branded installer shows resolved/prepared/installed dependency counts, approximate downloaded/total and remaining bytes, rolling transfer rate and ETA when uv exposes byte totals; carriage-return progress and ANSI output are normalized. The first uncached Windows attempt installed the full isolated runtime and 3.04 GiB model cache but exposed a stale model-readiness race after Enter studio; onboarding now awaits fresh catalogue and engine queries before dispatching automatic first sound. Three parser and two setup-gate regressions, locale parity and Electron typecheck pass.
- Live model lifecycle verification cancelled a running 5 GB VoxCPM2 download, installed and deleted a disposable 44 MB Zipformer model, translated one installed NLLB segment on CUDA, and unloaded that resident translator. No install or batch jobs, partial test models, or changed engine selections remained; the active TTS and dictation models stayed resident.
- Lazy workspace chunks now recover once from Vite's stale-preload event by refreshing the current route, with a session guard that exposes a repeat failure to the existing error boundary instead of looping. This covers the transient Gallery/Dubbing
Outdated Optimize Depcrash class after dependency changes. - Real profile verification generated a take, saved it through the Recent takes pane, preserved the active clone draft/settings, and removed its disposable take/profile. A second blank-transcript clone upload used the selected installed Faster-Whisper model locally, persisted a 94-character reference transcript, and cleaned the profile. The Dubbing workflow smoke now follows the compact rows' click-to-edit mechanic and again passes upload, streamed ASR, segment/timeline edits, translation, reload recovery, generation and exports.
- Windows native parity now covers folder selection and watch pause/resume/stop, the registered global dictation shortcut in hold/toggle modes, recorder pause/stop/cancel/reuse, and ordered insertion into a controlled external app with clipboard restoration. The run exposed two integration faults: the trusted recorder could not request its backend WebSocket URL, and clicking Stop attached foreground activation to the destination thread instead of the current foreground thread. Both are fixed. Linux Wayland tray/watch/clipboard fallback now pass; macOS permissions, Linux portal shortcuts, physical microphones and real target insertion remain native release gates.
- Profile action parity is verified through the current Profiles library rather than the retired Clone editor injection. Hovered Preview/Edit/Delete controls own their hit area above the row selector, and a live profile completed visible Edit, full and preview-only persona export, backend bundle inspection, Gallery re-import and cleanup. Model management and profile failure-recovery checks pass alongside the existing transcription workflow smoke.
- Transcription action parity now has an installed-model browser proof: Electron selected Accurate, uploaded the bundled conversational WAV through the visible file control, received timed Faster-Whisper segments, persisted and exported the transcript, handed its text and language to Voice cloning, then deleted the history entry. The isolated browser left user history and model selections unchanged.
- Model-selection parity now has a real UI round-trip: Settings switched ASR from in-process Faster-Whisper to its crash-isolated runtime and back, Engine Ready's live hover details named each selected runtime, and the backend retained
deepdml/faster-whisper-large-v3-turbo-ct2throughout. The verifier restores the exact original backend and checkpoint even after a failed assertion and starts no downloads. - A clean Ubuntu 26.04 WSL workspace now installs, builds and packages the Linux Electron shell without relying on Windows' transitive dependency tree. The resulting
linux-unpackedapp reaches explicit first-run consent, connects to a live same-origin backend, bundles the branded assets and native desktop helper, and passes packaged Wayland tray recorder, watch lifecycle, clipboard-only delivery and 150% DPR/no-horizontal-overflow checks. WSLg has no desktop portal service, so portal global-shortcut registration, a physical microphone, signed package delivery, updater behavior and macOS remain native release gates. - Linux distribution artifacts now build as a 240 MB AppImage and 182 MB Debian package with the shared icon,
VoiceStudiostartup WM class and package maintainer metadata. The AppImage and an actual Debian install both launched under WSLg Wayland and answered the bundled native-helper protocol. An isolated AppImage profile completed explicit runtime installation from the dependency cache, reached backend 0.5.2, relaunched through runtime reuse, and retired its backend run sentinel on both deliberate exits; the run exposed and fixed the missing POSIX nested-operation ownership descriptor. The AppImage also passes isolated tray capture and watch-folder pause/resume/stop with its real helper. Signing, updater delivery, a fully uncached package-to-first-sound run and a non-WSLg desktop remain release gates. - Runtime installation now keeps VoiceStudio branding in the fixed native title row while progress and logs scroll independently below it. The Windows packaged first-run smoke resized to the 960×640 application minimum and verified the complete header remained pinned in view, its scroll region began below the header, the layout had no horizontal overflow, and no installation started without explicit consent.
- Electron release packaging now validates the exact platform feed name, branded artifact name, frontend-derived version, byte size, SHA-512 and Windows differential blockmap before upload. A post-matrix release contract also verifies all four published manifests, payloads and GitHub digests against the stable or stamped Preview version. Preview checks compare both channel manifests and use the newer release while preserving the user's Preview preference and rejecting downgrades. The local stable Windows NSIS manifest matched its 202,238,435-byte installer, and both the packaged app and installer expose the exact shared 32×32 VoiceStudio icon rather than Electron's default. A published Stable/Preview update remains the delivery gate.
- The checkout-free repair path now has a native packaged regression: a clean profile exposed no source path, retained enabled Diagnose/Fix controls for an explicit
ACTION_REQUEST, rejected an ordinary diagnostic request at Electron main, and advanced an app request to agent availability rather than asking for a repository. No agent was launched and no backend mutation occurred. - Checkout-free repair prompts and completion UI now stay in app-operation language: the agent must verify the requested runtime state and report only remaining user input, while branch review and PR guidance appear only when a writable source checkout was attached.
- Native backend-crash restoration is now exercised through the real supervisor instead of only filesystem units. An isolated owned process exited with code 23, remained in the intended replacement-backend grace period, then rendered its log and exit metadata, persisted the evidence across an Electron relaunch, and excluded the later deliberate quit from the journal. The smoke now waits beyond the product's ten-second handoff window and cleans its isolated profile.
- Bug reports refresh Electron's native backend status before composing diagnostics, so the durable crash journal, exit metadata and bounded log tail remain available even while every Python API request is offline. A focused renderer regression covers the backend-down path.
- Checkout-free Claude Code and OpenCode runs now receive only a temporary VoiceStudio MCP API tool with default-deny permissions; no shell or filesystem tool is exposed. Packaged Windows read-only acceptance passes for Codex, Claude Code and OpenCode against the live OmniVoice CUDA runtime. Pi was unavailable on the host and remains an explicit acceptance gate.
- Runtime setup now uses the current pinned uv 0.12.13 across Electron, Tauri and release bundles. Electron packages the same per-platform executable already fetched for Tauri, removing the uv download from first launch. Its verified download/Python caches live beside the replaceable project, so Clean & Retry rebuilds the environment without discarding several gigabytes. Progress clears the last artifact as soon as announced downloads finish and explicitly reports package installation instead of looking stuck.
- The unpacked Linux package completed explicit first run under Ubuntu 26.04 WSLg Wayland from empty dependency and Hugging Face caches with bundled uv 0.12.13. Preflight rejected the constrained WSL model disk, accepted a host-backed cache, installed the 3,267,470,260-byte OmniVoice model, loaded CUDA on the RTX 4090, produced a paused 341,804-byte WAV, and shut its managed backend down cleanly. macOS remains the clean installer-to-first-sound release gate.
- Repair-agent discovery now rejects cross-OS executable shims. A packaged Ubuntu 26.04 run no longer advertises the Windows OpenCode npm shim inherited through WSL PATH, while native MCP-only OpenCode and Claude prompts identify their exact scoped tool without asking for shell or environment-file access. Native Linux agent acceptance remains a release gate.
- Profiles Gallery now has live remote-recovery proof: the Community catalogue survives a forced outage, retries in place, then completes real favorite, preview and Designer handoff actions without a renderer error. Its standalone acceptance explicitly bypasses first-run setup instead of depending on stale browser storage.
- Packaged Ubuntu profile parity now covers real profile creation, native
.ovsvoiceSave As, backend bundle inspection and cleanup using a reused installed runtime. The acceptance uses isolated app/backend data, handles the legitimate first-run analytics choice, and starts no dependency or model download. - Native renderer-generated Save As now passes through both Windows Electron and the packaged Ubuntu 26.04 application. Combined with the longform browser byte checks and real M4B render, this closes Linux export plumbing; macOS remains the native longform export gate.
- Closing packaged Windows VoiceStudio during “Preparing GPU libraries” now reaches a tiny authorized shutdown signal before the deferred system router exists, retires the run sentinel, and then permits Electron's bounded process-tree stop. A real packaged early-quit acceptance leaves no sentinel, preventing the next launch from reporting a false backend crash.
- Saved-profile portrait search now has live Electron acceptance: an explicit user click returned five labeled choices through the no-key Google/Openverse path, persisted one as a normalized JPEG, refreshed the editor avatar, and removed the disposable profile.
- Packaged Windows TTS status now retains the identity of the model that actually became resident. The live backend reported
k2-fsa/OmniVoice, a non-empty UTC load timestamp,cuda:0, and 1937.2 MB allocated VRAM; renderer refreshes can no longer turn that ready model into an anonymous checkpoint. - Stories now has a current real-model browser proof: the visible cast controls assigned an existing designed narrator and cloned guest to separate lines, the submitted render retained both profile IDs, and the completed result played through the shared media controls. The acceptance harness now crosses first-run state explicitly and follows the current
Add First Linemechanic. - Required CI now runs the current Profiles-to-Clone Vidstack acceptance and complete Dubbing workflow against a deliberately unreachable backend. Their setup, engine, model and compute-target fixtures are self-contained, so Profiles routing, paused generation completion, seek/exclusive preview behavior, Agent translation and Dubbing generation cannot pass by borrowing a developer's live runtime.
- Electron now consumes the backend's real-time event stream instead of leaving profiles, takes, projects, exports and model state stale between local or remote mutations. Packaged authenticated connections mint a path-bound
/ws/eventsticket, local development waits for health before opening its Vite proxy, and reconnect/unmount behavior has focused regression coverage. - The maintained live-route acceptance now follows the production hash-router mechanic through all 37 concrete workspaces and Settings views against the real backend. It fails on route fallbacks, missing headings, uncaught exceptions, console errors or warnings, request-load failures or HTTP errors. The current Windows development run passed cleanly; the transport contract documents why legacy settings and MCP calls intentionally appear as
/api/api/...before the desktop proxy removes its prefix. - The required backend route inventory was regenerated from the live application after the parity audit exposed 19 unsnapshotted endpoints. It now guards the complete performance-profile, model-install/access, profile-image, diarisation/audio.cpp, caption, Agent Fit, translation-pack, Batch retry and clean-shutdown contracts instead of allowing those Electron/Tauri workflows to drift outside the API gate.
- Live-operation bug reports now attach a bounded, scrubbed tail from the current backend log in both desktop shells. Engine failures that leave Python healthy, including the macOS MPS generation reports that exposed this gap, now arrive with the relevant backend evidence instead of only a generic
RuntimeError; report creation remains explicit and review-before-submit. - OmniVoice's isolated MPS runtime now heartbeats during silent cached model loading, serializes concurrent progress frames, logs the bounded child traceback, and reaps an unusable protocol pipe before releasing the request lock. The fail-before recovery race and shared sidecar lifecycle suite pass; the running Windows app restarted cleanly with zero renderer errors, while physical Apple Silicon generation remains a release gate.
- Live acceptance is derived against all 37 rendered routes (plus the
/settingsredirect in layout coverage); a parity-inventory regression now fails when router, model-family and real-backend route coverage drift. - Asynchronous renderer errors outside React boundaries now enter the same scrubbed, opt-in automatic Repair Agent path as route failures. One shared classifier keeps Abort cancellation, injected performance-monitor faults, extension top frames and duplicate events out of both repair and the frontend error count without hiding an app fault that merely has an extension deeper in its stack. Electron startup promises terminate through the durable main-process journal instead of escaping as unhandled rejections.
- All nine secondary workspaces now use the same compact stacking contract as Dubbing. At a 640 px CSS viewport, each local pane remains bounded above a primary workspace at least 580 px wide and 480 px tall; Transcribe no longer lets an empty History pane push its upload, recording and dictation actions off-screen.