Skip to content

VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.

16.8 KB
Raw Blame

Installer review: remaining findings and overlap

Source-review baseline: fb9a960c88744edc36daf2cf38d48df0971ee713, 2026-10-04, app 0.5.6. Proposal; no finding is fixed by this docs-only patch. The active renderer is electron/src/renderer/src, not the legacy shared wizard. Historical tests below used a clean archive of that commit, not the integrated branch.

Integration refresh: 2026-10-05

The branch now includes main 990f0627ba14dbb309b88c0051ebb953d3101800, including the CPU-budget change from #2611. This proposal adds no runtime behavior. This is a documentation integration refresh, not a rerun of the source probes or native acceptance. The findings, line-pinned evidence and local results below remain the 2026-10-04 snapshot and must be revalidated before implementation.

  • #2602 merged at 5258077d5967d860da3954bad11be447256e8a61, including disk-headroom wording, Rosetta guidance and runtime package-progress identity. Preserve these fixes; they are no longer open-PR work
  • #2605 merged at 3fcd381e5c3282b084384dbc38be28ddaba0e6bf, including source CPU selection/no-sync restarts and native compatibility helpers
  • #2585 remains open at e87cba0449485b226674961be796b9d3e20ed0a1. Main has absorbed overlapping fixes through #2602; compare the remaining diff before using this branch as implementation evidence
  • #2607 remains a draft, now at 51e4eb33cc7cbec39493cfeb9027e6a6b15710fb. Its older reviewed head below is historical. Reconcile the current Windows ROCm work with the landed #2605 selector/interpreter changes and retain its native qualification gates

These status checks do not establish new platform support or merge readiness. Fresh hosted CI on the integrated PR head is the merge gate; none of the local test or hardware results below is claimed for that head.

Already addressed at the source-review baseline

  • #2500: packaged non-NVIDIA x64 CPU wheels, CPU float32 and experimental Windows ARM packaging. Do not reintroduce the old broad CPU/CUDA bootstrap claim
  • #2578: Electron-native platform install guides, AppImage/deb truth, current uninstall folders/elevation safeguards, best-effort .pth repair and stable-only desktop updates. Retire old finding R11 (legacy install guides); do not propose restoring Preview or blanket uninstall-script changes
  • #2557: package-managed installations can disable in-app updating via VOICESTUDIO_DISABLE_UPDATER=1; the Linux guide covers the community AUR package

These merged changes do not establish native acceptance for this review.

Open-PR overlap

Historical snapshot: read-only PR metadata/diffs were checked on 2026-10-04. For current merged/open status, use the integration refresh:

PR / reviewed head Relevant work already proposed Treatment here
#2602 · 4cf3db00993b910a50c00b19ca6cb3cb823b440b · open Named disk-headroom/total warnings, Mac architecture/Rosetta guidance and runtime package-progress identity Do not implement these again; recheck after landing. Headroom wording is distinct from R5/R7; runtime package bytes are distinct from R3's model-pack aggregation
#2585 · e87cba0449485b226674961be796b9d3e20ed0a1 · open Runtime package-progress identity, storage scan/cleanup and broader desktop/audio fixes Avoid duplicate progress/storage work. Its general storage cleanup changes are distinct from R8's sidecar uninstall function
#2605 · 441673e1ff5423e6713aef834ec0b1b62dc6c703 · draft Source CPU/CUDA dependency groups, CPU selection before sync, no-sync restarts, extracted native-compatibility helpers and Electron lazy-executable startup Reuse this source-install work; packaged CPU bootstrap already landed in #2500. R6's remaining recipe edges are not resolved by this CPU change
#2607 · 265295e7e1b1687308314aefe27e219c5994e95f · draft Broader Windows AMD/ROCm desktop and engine integration, depending on #2600's source recipe; Windows x64 Python 3.12, HIP Torch/CT2 checks and recipe-aware runtime identity Reconcile with #2605 and #2602 instead of creating another selector or recipe. Keep the documented separate source flow until launcher/interpreter selection agrees; this draft does not establish shipped or qualified AMD support

No open-PR tests or merge readiness are asserted here. Their scopes overlap each other too; this proposal does not pick a winner or absorb these branches. Refresh this snapshot before implementation, especially if any reviewed head changes.

Historical #2605 / #2607 reconciliation: seven paths overlapped at the heads above: CHANGELOG.md, docs/install/windows.md, electron/src/main/runtime-project.ts, electron/src/main/runtime-project.test.ts, electron/src/main/runtime-torch-variant.test.ts, scripts/setup.py and tests/test_setup_rocm_variant.py. Preserve #2605's helper extraction and optional cuDNN HIP exclusion alongside #2607's recipe-aware runtime work. Their runtime tests independently add global fetch mocks (success versus offline rejection); choose one intentional baseline with per-test overrides, not duplicate stubs. The source selector in #2605 falls back to CPU/CUDA for a Windows ROCm request, and its launcher does not select Python 3.12 on x64. #2607's setup gate requires Windows x64 Python 3.12: combining the branches unchanged can sync a CPU/CUDA stack first, then reject an existing 3.11 environment. Preserve its separate source recipe or reconcile and test the launcher before presenting a unified path. #2607 also retains clean Windows/packaged-artifact, engine-coverage and non-ASCII-path validation gates; its reported hardware experiments are not acceptance for this proposal.

Remaining findings

P1: correctness reaching working first use. P2: recovery/resource/usability gaps. Priorities do not imply observed data loss or a measured user failure rate.

ID Evidence / priority Remaining behavior at the baseline Smallest next step
R1 Isolated source probe · P1 Setup's positive-size cache predicate disagrees with catalogue completeness Reuse a target-aware completeness/readiness projection
R2 Source review · P2 Simple pack view excludes failed rows and has no retained failure/Cancel controls; detail requires Advanced Expose existing recovery in the simple view
R3 Isolated source probe · P2 Pack percentage sums only active jobs, so completed work disappears Retain completed work in the accepted-plan denominator
R4 Source review · P2 Model/sidecar state is process-local; detailed model failures disappear from status after 60 seconds Separate retry cooldown from terminal history; durable interrupted state
R5 Source review · P2 Pack confirmation uses catalogue estimates; exact file/dependency metadata resolves inside the started job Read-only plan before confirmation; explicit size uncertainty
R6 Source review · P2 Explicit ROCm syncs default wheels before replacement; explicit CUDA/default precedes the Windows ARM automatic CPU branch Validate recipes and avoid CUDA-first replacement; reconcile #2605 and #2607 before adding runtime work
R7 Isolated source probes · P2 Both backend guards treat measured zero free bytes like a failed/unknown probe Represent unknown separately; reject positive work at zero
R8 Isolated source probe · P2 Managed sidecar removal ignores deletion errors and clears registration despite remaining files Observe removal outcome and preserve retry/registration on failure
R9 Source review · P2 Wizard position and preview selection are component state; persisted setup flags cannot restore them Persist and reconcile a versioned setup session
R10 Source review · P2 Most preflight text bypasses i18n; compact progress has no progressbar semantics; consent details targets an absent FAQ Localized reason codes, accessible status/focus and valid privacy details

Evidence details

R1: setup_status calls is_cached, while the catalogue also calls cache_is_complete. The setup gate uses models_ready to advance. A fresh AST-extracted probe ran the unchanged predicate bodies with a temporary snapshot containing only a 25-byte config file and mocked HF inventory: is_cached=true, cache_is_complete=false, models_ready=true, missing=[]. This proves predicate disagreement, not a native click-through or a full completeness specification.

R2–R3: PerformanceModelPacks filters all terminal states before summing bytes. Executing those exact aggregation statements with two 100-byte jobs at 90 and 10 bytes gives 50%; completing the first at 100 gives 10%, though the combined total should be 55%. A terminal-only set has zero visible active rows. Detailed recovery exists in the same module; this is not a claim that the app has no cancellation or retry implementation.

R4: Model job status omits failures once now - failed_at >= 60; its maps and the sidecar jobs dictionary are in memory. Existing range manifests, cached partial files, runtime markers and renderer reconnection remain valuable, but do not provide durable user-visible operation history.

R5: The pack resolver sums size_gb. The download worker resolves pinned snapshots/dependencies and checks space before weight transfer, but after the user has started the job. No concrete size undercount or transfer benchmark was measured. #2602's headroom warning does not introduce a pre-confirmation plan or combined per-volume reservation.

R6: Runtime selection/sync excludes CUDA packages for CPU only, then performs default frozen sync before the ROCm replacement. Explicit cuda/default returns before the Windows ARM auto-CPU branch; the ROCm platform gate is OS-only. This is source evidence of unsupported/unverified recipe edges, not proof of native ARM/ROCm installation failure. The shipped auto-CPU path is already present; avoid a second implementation while CPU/runtime work is active. #2605 preserves these baseline edges while adding source CPU selection. #2607 proposes a narrower Windows x64 ROCm recipe and changes runtime selection; it must be reconciled with #2605, not treated as proof that baseline R6 is fixed. Recheck default-sync ordering and explicit Windows ARM overrides in the combined code, with separate source and packaged acceptance.

R7: The model guard and sidecar guard both return no error for free <= 0. Fresh extracted-function probes with positive requested bytes and zero measured free reproduced this for both. This concerns direct backend admission; the compact UI's low-disk gate can still block its own path. Existing preflight and mid-transfer ENOSPC handling are not claimed absent.

R8: Sidecar uninstall uses rmtree(ignore_errors=True) and bases success on pre-deletion existence. A fresh unchanged-function probe with an isolated managed directory and a mocked ignored deletion failure returned uninstalled, cleared environment/preferences and left the file present. No real Windows sharing violation was exercised. #2578's app-uninstall script safeguards and open storage-temp cleanup work are different paths.

R9–R10: Setup flags store started/completed only; the gate initializes step zero. Preflight translates only the architecture case. Compact pack progress uses plain divs. Consent correctly saves enabled:false for No thanks and waits for success with equal outline buttons; its details URL points to README#-faq, absent from the baseline README. There is no newly reproduced consent reversal, full-Max selection bug or performance benchmark. Source inspection is not a screen-reader/viewport audit.

Historical verification: 2026-10-04

Check Result / limits
sh -n scripts/install.sh Passed
From electron/: node --test tests/shell-installer.test.mjs tests/installer-worker.test.mjs tests/native-linux-libraries.test.mjs 27 passed, 0 skipped; Linux / Node 24.19.0; fixture contracts, not native installers
Python 3.12.14 / Node source probes Reproduced R1, R3, both R7 guards and R8 without ML dependencies/network
Local Markdown targets/anchors, pinned source paths/line ranges, SVG XML, git diff --check, clean-baseline patch application Passed
SVG Existing illustrative asset rendered and visually inspected; not a product screenshot

No runtime/dependency/workflow/version files changed. The previous proposal's 19-pytest result is historical and is not carried forward as a fresh pass: pytest is unavailable in this review environment and no software was installed.

Not run: full backend/React/Vitest suites, typecheck, builds, PowerShell, packaged install/first-output/update/rollback, native picker/accessibility tools, macOS/Windows/ARM/GPU hardware. Proposed designs and source-level defects are explicitly separate from platform acceptance.