Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
chore(release): complete the 1.10.1 changelog (#2198)
Chad Piha82fda520853e
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 01020fa03..72ad585fa 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -9,7 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0## [1.10.1] - 2026-09-14-A follow-up to 1.10.0 built from everything that landed since. Notes open straight onto their AI Summary when one exists and the raw tab becomes **Your notes**; the "Enhanced" label is gone in favour of **AI Summary** everywhere, a meeting that ends on its own still offers the summary, and the post-meeting notification is retired. Team spaces get one **Settings** dialog with a built-in emoji picker, the sidebar gains **Invite your team**, and the tray and dictation pill pick up quick actions. Uploads accept video and nearly every audio container, Windows finds the right microphone again and starts listening sooner, local dictation skips a decode step, and bring-your-own-key Deepgram and AssemblyAI work again. Cleanup no longer pastes a reply the model cut short, whisper.cpp stops dropping words at window boundaries, local models size their context to the request, and Reset app data restarts the app so history keeps working.+A follow-up to 1.10.0 built from everything that landed since. Notes open straight onto their AI Summary when one exists and the raw tab becomes **Your notes**; the "Enhanced" label is gone in favour of **AI Summary** everywhere, a meeting that ends on its own still offers the summary, and the post-meeting notification is retired. Team spaces get one **Settings** dialog with a built-in emoji picker and a single flat member roster, Teams are called **Groups** in workspace copy, the sidebar gains **Invite your team**, and the tray and dictation pill pick up quick actions. Uploads accept video and nearly every audio container, Windows finds the right microphone again and starts listening sooner, local dictation skips a decode step, and bring-your-own-key Deepgram and AssemblyAI work again. Cleanup no longer pastes a reply the model cut short, whisper.cpp stops dropping words at window boundaries, local models size their context to the request, assistant answers finally render tables, and Reset app data restarts the app so history keeps working.### Notes@@ -24,12 +24,13 @@ A follow-up to 1.10.0 built from everything that landed since. Notes open straig- **The recording timer survives switching notes.** The elapsed time lived inside the header control, and the editor is rebuilt for each note, so leaving a live recording and coming back restarted the clock at 00:00. The session's start time now lives in the meeting store and the header derives the elapsed time from it, which also stops the timer drifting while the window is throttled. (#2162)- **A meeting that ends on its own still offers Generate AI Summary.** The offer was gated on the Transcript tab, and every detected or quick-action meeting opens on your notes, so the one time a meeting ended without you the button was hidden. The offer no longer looks at the open tab. (#2162)-- **The post-meeting notification is gone.** The card that announced the meeting had ended and offered thirty seconds to restart the recording has been removed. Recordings still stop automatically when the call does and the transcript is saved exactly as before; if a stop fails to persist, the "Recording stopped automatically" toast remains. (#2192)+- **The post-meeting notification is gone.** The card that announced the meeting had ended and offered thirty seconds to restart the recording has been removed. Recordings still stop automatically when the call does and the transcript is saved exactly as before; if a stop fails to persist, the "Recording stopped automatically" toast remains. (#2192, #2197)- **Windows system-audio capture ignores a stopped helper's last words.** When a capture session was stopped or replaced, stderr output from the old helper process was still fed into the new session, corrupting its buffer and raising errors against the wrong recording. Late output from an inactive helper is now dropped, matching what the macOS capture manager already did. (#2017, thanks @hsusul)### Team spaces-- **One Settings dialog per space.** A space's menu now opens a single **Settings** dialog with two tabs. **General** holds the name-and-icon field, **Leave space** for members who hold access through a team, and **Delete space** for admins behind the existing type-the-name confirmation. **Members** is the teams-and-members roster: each row has one quiet menu with **Admin**, **Member** and **Remove**, and **Add people** stays collapsed until you search. Opening members from a note's header lands in the same dialog. Names save on blur, Enter or close, and a draft you never touched can no longer overwrite a teammate's rename. The old inline rename, the separate members dialog and the **Change emoji** item are gone. (#2186)+- **One Settings dialog per space.** A space's menu now opens a single **Settings** dialog with two tabs. **General** holds the name-and-icon field, **Leave space** for members who hold access through a group, and **Delete space** for admins behind the existing type-the-name confirmation. **Members** holds the roster described below. Opening members from a note's header lands in the same dialog. Names save on blur, Enter or close, and a draft you never touched can no longer overwrite a teammate's rename. The old inline rename, the separate members dialog and the **Change emoji** item are gone. (#2186)+- **Members is one flat list of people.** Space settings → **Members** now shows everyone with access in a single roster whether they were added directly or through a group, with rows marked _via Design_ where a group is what grants it. Each row carries one quiet menu — **Admin** and **Member** with a line explaining each, then **Remove** for direct members; a role inherited from a group cannot be lowered here and the menu says so. **Add people** searches the workspace roster and adds someone directly, and typing an unknown email invites them to the workspace and the space in one step. Groups move to a collapsed **Groups with access** section keeping the existing assign, access-cap and unassign controls. Creating a space no longer requires a group: name and icon, then people and groups if you want them, and whoever creates it becomes its admin. Leaving drops only your direct membership and tells you when a group still grants access; removing a direct member says the same. Teams are called **Groups** in workspace copy across all 11 locales — "Teamspaces" and copy about the human team are unchanged. (#2194)- **A built-in emoji picker replaces the OS panel.** The system emoji panel behaved differently on every platform and never worked on Linux, which only ever saw a fallback grid. The new picker is searchable and categorised, with keyboard navigation, recents and a **Remove** action, and its labels and search terms are translated into ten languages from emojibase data. Emoji the OS font cannot draw are hidden after a one-time check; country flags are hidden on Windows, which has none. (#2186)- **Invite your team from the sidebar.** Workspace owners and admins get an **Invite your team** row above **Get a free month** that opens the existing invitation dialog for their workspace. Members, signed-out users and users with no workspace do not see it, since the invitation call requires the admin role. (#2181)- **An existing Screen Recording grant counts as the screen-context opt-in.** Resetting onboarding on macOS no longer shows the row as disabled while the OS permission is already granted. Windows keeps its explicit Enable. (#2186)@@ -40,7 +41,9 @@ A follow-up to 1.10.0 built from everything that landed since. Notes open straig- **Dictation starts listening sooner on Windows.** Pressing the hotkey could take two seconds before the microphone opened. Since 1.9.0 the app pins the system-default microphone to a real device so its health checks can run, and on Windows finding that device means starting a PowerShell process that compiles C# on the fly — a lookup that ran on the first dictation after launch and again whenever an audio device came or went, and expired after 30 seconds either way. The answer is now resolved once at launch, kept until the OS reports a device change, and shared between windows, so the hotkey path only has to open the microphone. (#2178)- **Learned corrections skip everyday words.** When you fixed a transcript by swapping one common word for another, such as "why" for "what", the app treated the replacement as a mishearing and saved it to your dictionary. Those entries taught the cleanup model nothing and padded every request. Corrections that resolve to a common English word are no longer learned; names and jargon are unaffected. (#2178)- **The dictation pill stops jumping while you talk.** Every live-transcript chunk asked the window to show itself and re-enforce always-on-top, which on Windows and Linux meant a ShowWindow and two SetWindowPos calls against an already visible pill, so the preview flickered and redrew as you spoke. Both calls now run only when the pill is actually hidden, the idiom the other panels already used. (#2185, ported from #1995 by @hsusul)+- **The Linux dictation pill stays hoverable.** The cancel **X** appeared on the first hover and never again, so a recording could no longer be cancelled from the pill. Returning the window to click-through asks Electron to forward hover events, which only macOS honours; on Linux it left a 1×1 input region that nothing could hover back into. Linux now keeps normal hit-testing, as Windows already did, and the pill also stays visible after **Retry**. (#2161, thanks @boseq)- **A cleanup reply the model cut short is no longer pasted.** When cleanup hit its output-token limit, the truncated text was pasted as if it were complete, so a long dictation could end mid-sentence with no warning. Cleanup now requires a complete reply: a truncated or empty one keeps your raw transcript, pastes that instead, and raises the **Cleanup failed** toast with a translated reason. Retrying cleanup from history reports the failure rather than saving the partial text, and Gemini's output cap is removed so a heavy-thinking model cannot trip it. (#2130)+- **Gemini cleanup no longer pastes a half-finished dictation.** When the cleanup response stopped early — Gemini counts its hidden thinking against the same output budget, so a long dictation could exhaust it — the truncated text was pasted as though it were the whole transcript, with nothing to say a sentence had gone missing. An incomplete or empty response is now rejected so the original transcript survives through the existing fallback. Output headroom was raised, thinking controls are set per model, Gemini 3 uses Google's recommended default temperature unless you override it, and the cleanup instructions are sent separately from your words. Other providers and local models are unchanged. (#2092, thanks @IdrisGit)- **Snippets whose trigger starts with the assistant's name expand instead of waking it.** A trigger such as `openwhispr review` was read as an address to the assistant, which stripped the name and sent "review" off as a command. Any match for the assistant's name that falls inside a snippet trigger is now ignored, in both detection and the address stripping, so the two can never disagree. A trigger that is exactly the bare name still shadows the wake word. (#2153)- **Quick actions in the tray and the pill.** The tray menu now leads with **Start listening** / **Stop listening**, **Ask assistant** and **Start meeting recording**, and the dictation pill's right-click menu gains **Start meeting recording** while idle. Each action runs where it belongs: listening goes to the existing dictation start and stop, a meeting goes through the same entry the meeting hotkey uses and brings a live recording forward instead of opening an empty note over it, and only **Ask assistant** asks the renderer, since only it can open the panel. If organisation policy forbids the assistant, or policy is still loading, the pill surfaces and says why. (#2163)- **Windows: a dictation started from the tray pastes where you were working.** Starting from the tray captured OpenWhispr's own menu-owner window as the paste target, so the fast-paste helper pulled the foreground onto OpenWhispr, dropped the keystroke and restored the clipboard over your text. A target that is ours but is not one of our real windows is now dropped; the note editor remains a valid target. (#2163)@@ -73,7 +76,9 @@ A follow-up to 1.10.0 built from everything that landed since. Notes open straig### Assistant-- **Voice Assistant on OpenWhispr Cloud with a stray "Separate vision model" switch.** On a managed workspace, an override that was switched on but never given a model was shown with an invented provider and model, and every screenshot command went to that provider without a key, failing with "OpenAI API key not configured". Workspace policy no longer invents a target for an override the user never configured, and a chosen override that policy moves to another provider is cleared rather than repointed, so it asks to be picked again. A failed assistant request is now written to the debug log.+- **Voice Assistant on OpenWhispr Cloud with a stray "Separate vision model" switch.** On a managed workspace, an override that was switched on but never given a model was shown with an invented provider and model, and every screenshot command went to that provider without a key, failing with "OpenAI API key not configured". Workspace policy no longer invents a target for an override the user never configured, and a chosen override that policy moves to another provider is cleared rather than repointed, so it asks to be picked again. A failed assistant request is now written to the debug log. (#2139)+- **Assistant answers render tables.** A table in an assistant reply collapsed into a single paragraph of literal pipes, and always had: the renderer follows CommonMark, which has no table syntax. Headings, bold and lists were never affected. (#2131)+- **Long assistant answers stream more smoothly.** Every streamed token re-parsed the entire answer as markdown, so a long reply got slower the further it ran. Streamed text now reaches the screen at most every 32 ms, flushed immediately on completion, cancel or close, so the answer you end up with is unchanged. (#2171)## [1.10.0] - 2026-09-11@@ -115,7 +120,6 @@ The desktop app gets a new look: an inset content container with a top bar and### Meetings-- **No more post-meeting card.** When a meeting recording stops on its own, the 30-second **Restart recording** card and its fallback toasts no longer appear. The recording still stops and the transcript is saved, and the note offers **Generate AI Summary** as before. The card's main-process, overlay, preload and translation code has been removed with it. (#2192, #2197)- **Recording works again after a reload.** A renderer reload mid-recording left the main process holding the session, so every later start failed with "Operation in progress" until the app restarted. The session now stops when its window navigates away, and a restart from the same window retires the stale session first.- **Smoother notification swipe.** The meeting notification card follows your pointer, fades as it goes, and leaves through the edge you dragged it toward. The window no longer shows scrollbars mid-animation, and on macOS it now sits above every other window, including OpenWhispr's own floating panels.- **Meeting recordings that went silent partway through on macOS.** The macOS audio tap is built once against the devices present when recording starts and never follows the machine afterwards, so a route change — headphones connecting, a device going away — could stop capture while the helper process stayed alive. Nothing noticed: no exit, no error, and silence from a dead tap is indistinguishable from a quiet call, so the rest of the meeting was lost. A watchdog now restarts the helper when no audio arrives for 6 seconds or the pinned device reports that it changed, up to three attempts, and tells you either way. A genuinely quiet call only warns, since restarting a working capture would cost real audio. (#2040, thanks @alekc)