Changelog
Written for the person deciding whether to upgrade, not for the person who wrote the commit. Parsed at build time from the repository’s own CHANGELOG.md, so it cannot drift from it.
Fixed
- Dictation works after copying a screenshot: preserve the image when borrowing the clipboard, instead of refusing to paste.
- Use paced typing when clipboard formats cannot be safely preserved, while respecting cancellation and the original target window.
- Return the borrowed clipboard after both successful and failed delivery, without replacing anything copied in the meantime.
- Recovered transcriptions appear immediately in Home and wait for an explicit Copy. Dictation no longer leaves a 15-second clipboard backup or copies retry results automatically.
- Cancelling hands-free dictation no longer consumes the next hotkey press. Recording limits, microphone failures and shortcut changes also reset hands-free and click-to-talk controls.
- Starting hands-free dictation preserves earlier speech still processing. Rapid cancel/restart no longer lets an old stop request close the new recording.
- Retrying older audio after a newer successful dictation keeps the recovered text available for Copy.
Fixed
- Update TLS and certificate-store dependencies to address current security advisories.
- Preserve signed executable updates while removing unused ZIP update extraction.
- Upgrade the native build script's ZIP dependency without changing the speech runtime or models.
Validation
- Broader speech accuracy, application latency, model/GPU comparisons and live hardware acceptance remain unverified.
Fixed
- Automatic paste no longer rejects its own clipboard write after Windows finalizes text formats.
- Preserve meaningful dictated words and quiet speech; match saved corrections through the formatter's token rules.
- Guard delivery with the original window and cancellation, preserve intervening clipboard copies, and retain complete text for explicit Copy when delivery is held.
- Synchronize settings revisions and save errors; harden downloads, transactional history migration, local-day streaks and running retention.
Added
- Restartable local decoding and bounded, memory-only failed-audio retry.
- Strict real-asset accuracy and end-to-end latency gates that reject missing or unapproved inputs.
Validation
- Native unit tests, real v2 model smoke tests, controlled decoder recovery and bundled installer payload checks pass.
- Broader speech accuracy, thirty-dictation application latency, alternate model/GPU comparisons and live hardware acceptance remain unverified.
Coding-tool names and git commands are spelled correctly far more often.
Fixed
- "Claude Code", "Codex", "Docling", "PaddleOCR", "OpenCV", "Vercel" and "Tauri" used to come out as "clawed code", "dockling", "Paddle OCR", "Open Tv", "Versal" and "Tori". On a test of 16 spoken sentences the error rate on those names fell from 14% to 2.5%, and everyday English was unchanged. The speech model is now told your Dictionary terms before it listens, so a word you add helps the next dictation, not just the clean-up afterwards. Silence and noise still produce no text. Both Parakeet models get this; the Whisper model does not.
- "get commit", "get push", "get pull", "get diff", "get checkout", "get rebase", "get stash" and "get clone" now come out as the git command. "get status" is ordinary English, so it is only rewritten in a terminal; add it to your Dictionary if you want it everywhere.
Changed
ov --hint is replaced by ov --no-hotwords. The vocabulary is offered to the speech model by default, as in the app; the flag turns it off to compare.
Changed
- The Flow Bar has a new look that matches the redesigned Hub in every theme. It is dark glass, and at rest it shows only the OpenVoice logo's seven bars: no text, no colour, nothing moving. The moment the microphone opens, colour arrives. The waveform runs from the logo's green through mint and cyan to periwinkle, your voice lights the glass from inside, and a thin line of light runs around its edge. When your words land, one pass of light goes round the rim.
- "Hold Right Ctrl" is no longer written on the bar. Your shortcut is in the bar's tooltip and next to Start dictating in its menu (right-click the bar).
- Messages on the bar carry an icon that says what happened. Text left on your clipboard is no longer shown in warning amber; only a real failure is marked, in coral. Downloading the speech model shows a progress line along the bar's bottom edge.
- The bar's menu has the same dark glass and an icon on every row, and it now fits in full. Before, a few pixels at its outer edge were cut off.
- While you dictate, the bar no longer blocks clicks in a 22-pixel band around itself. That band held a glow the new design does not use.
Fixed
- Screens taller than the window can be scrolled. The Advanced screen was missing its Files and Engine cards and Settings was missing the bottom of Privacy — including the button that opens your data folder — with no scrollbar and no way to reach any of it. Every screen now scrolls when it needs to, at any window size.
- Advanced opens on the rules that actually rewrote your sentence instead of listing all eight stages, four of which usually do nothing. "Show all 8 stages" opens the full pipeline when you want it, and a sentence no rule touched now says so.
- The Writing style screen showed a sample sentence the formatter would not produce: "so" is never removed as a filler, at either setting.
Fixed
- The Hub opens on its proper dark background. It used to open a pale, washed-out grey and stay that way until you resized it, minimized it or clicked around, which could be half a minute or more. The page itself was right; the window underneath it was never cleared at startup.
Fixed
- Dictionary, Writing style, Speech model, Settings and Advanced no longer stay on grey placeholder bars after launch. The Hub could ask for your settings a moment before the app was ready to answer, and it never asked again, so every screen but Home looked empty until you restarted. It now can't ask too early, and it tries again if a load ever fails.
- The Graphite theme opens onto a visible slate-and-green backdrop instead of flat grey. Its colours sat so close to the base that under Windows' Mica the window looked like it had not finished painting.
The Hub is rebuilt. The last thing you said is now the first thing you see, and the whole window is a piece of frosted glass over your wallpaper.
Added
- Your last dictation sits at the top of Home, in full, with Copy next to it and the app it went to. If it did not land, that card turns amber and tells you where the text actually is.
- Three themes, each in light and dark: Glacier, Graphite and Lagoon, in Settings > Appearance. Glacier is the default.
- Windows 11 Mica. The Hub is transparent and tints your wallpaper. On Windows 10, or with Transparency effects turned off, it paints its own quiet backdrop instead, and the panels go solid.
- Ctrl K opens a command palette that searches everything you have ever dictated and jumps to any screen. Ctrl 1 to Ctrl 6 go straight to a section.
- A History screen with day headings, search, and filters for Code, Terminal and Messages. Filtering and searching now happen in the database, so a dictation from months ago is findable; before, only the last 200 were.
- Fix a word on any dictation shows what OpenVoice actually *heard* and turns those words into a dictionary entry in two clicks.
- Home greets you by your first name, read from your Windows account. It never leaves the machine.
Changed
- Every Hub screen -- Dictionary, Writing style, Speech model, Settings, Advanced -- was rebuilt on the same set of controls, so they now look and behave as one app.
- Engine messages ("Nothing to paste yet") are toasts in the corner that leave by themselves, instead of a banner that pushed the whole screen down until you dismissed it.
- The Hub window opens at 1100x740 and no longer flashes white on a cold start: it is revealed once its page has loaded.
Removed
- Paste again on the Hub. It pasted into whichever window had focus, which was usually the Hub itself. Copy is the way to get text back out.
Fixed
- Home opens and resizes without stalling. The screen was building every one of the last two hundred dictations and then hiding the ones that did not fit, so it carried about 2,400 hidden elements to show fifteen rows. Dragging the window edge cost 320 ms of layout per step and returning to Home cost 184 ms; both are now 13-14 ms, and the window holds 86% fewer elements. Measured in the real app:
docs/benchmarks/2026-09-22-hub-ui-performance.md. - A dictation can no longer hang the app forever. If the speech model crashed mid-decode, that dictation -- and every dictation after it -- waited for a reply that was never coming. It now falls back to decoding the recording in one go, which is what every other failure already did.
- A speculative decode that the release did not use could keep running after your text had already been delivered, taking a core from whatever you were typing into.
Long dictations no longer look like they failed to paste in Claude Code.
Fixed
- A minute of dictation now shows up in full in Claude Code. Claude Code hides any paste over 800 characters behind a
[Pasted text #1] label, so a long paragraph seemed to vanish even though every word was there. OpenVoice now pastes long dictations in a few pieces, split between words, so the whole paragraph appears in the input box. Other apps receive exactly the same text as before, a fraction of a second later for very long dictations. - A long paste can no longer land half in one window and half in another. If you switch windows while a long dictation is being pasted, OpenVoice stops and leaves the whole dictation on your clipboard so you can paste it yourself.
Letting go of the key is now instant. OpenVoice starts turning your words into text while you are still speaking, so by the time you release there is usually nothing left to wait for, however long you talked.
Measured on the reference machine (Ryzen 5 6600H, CPU only), from release to text at the caret, in a first real session: dictations over fifteen seconds went from 2.2 seconds to 0.4 seconds typical, and short ones stayed around a quarter of a second. On recorded speech with a natural pause before release, most dictations have no decoding left at all. Accuracy did not drop; the full record is in docs/benchmarks/2026-09-13-release-latency.md.
Changed
- Dictation is decoded while you speak, so text appears almost as soon as you let go of the key, including after long dictations. It is cut at natural pauses and each finished part is decoded while you carry on; if anything about that goes wrong, the whole recording is decoded in one go exactly as before, so your words never depend on the fast path. (ADR 0012)
- Audio is converted to 16 kHz as it is recorded, and letting go no longer waits for the microphone to finish closing.
- After the app has sat idle, the speech model is woken as soon as you start speaking, so the first dictation of the morning is not the slow one.
Fixed
- Audio from accidental taps, silent recordings and cancelled dictations is no longer kept in memory until the app quits.
Added
ov latency shows where dictation time goes, stage by stage.ov bench measures release latency and word error rate on LibriSpeech.
A way back to the middle, and a bar whose edges survive being clipped.
Added
- Back to center. Right-clicking the Flow Bar now offers a way to put it back where it started. Dragging the bar somewhere awkward used to be a one-way trip unless you found the tray icon, which is not where anyone looks when the thing they want to move is on screen in front of them.
- The default resting place is now measured from the usable screen rather than from the glass. On a 1080p display with Windows 11's 48px taskbar the bar used to leave 48px of dead air beneath it; with the taskbar set to auto-hide it floated 96px too high, and docked to the left or top it was wrong the other way. It now sits a consistent gap above whatever the desktop actually ends at.
Fixed
- The Flow Bar's rounded ends are no longer pixelated. The window is clipped to the bar's silhouette, and that clip could only round corners to whole pixels — so it cut into the smooth edge underneath and left a visible staircase at both ends. It was worst on the yellow bar shown while the speech engine starts and when no audio is detected, because those are the only states that draw the outline in a colour bright enough to see it against the pill.
- The Flow Menu now closes when you click somewhere else. Right-clicking the bar opened a menu that nothing could dismiss but a second right-click on the bar itself — so a menu opened by accident sat over your work until you found your way back to it. It now closes on a click anywhere else, on Escape, when you switch to another application, when the bar is hidden, and after fifteen seconds of being left alone.
- The Flow Menu no longer opens on the wrong side of a bar that has just been moved, and no longer flips sides while you are reading it.
- A bar hidden with its menu open comes back as a bar rather than as a menu.
The Flow Bar gets out of your way when you are not using it — and stops disappearing when you are.
Added
- The Flow Bar puts itself away. After five seconds of quiet it folds down to a slim horizontal line; clicking that line, or starting to dictate, unfurls it again. It never folds away while it has something to tell you — a failure, a notice, or a session in progress all hold it open. Turn it off with "Stay full size" in the Flow Menu.
- Double-tap the dictation shortcut to leave the microphone open, so you can take your hand off the keyboard. Tap once more, click the bar, or press Escape to finish. A hands-free session is marked differently from a held one, because the difference is whether letting go stops it.
Fixed
- The Flow Bar could disappear and not come back until you restarted OpenVoice. Pressing the shortcut did nothing visible, and neither the folded line nor the full bar could be found anywhere on screen. The window was being clipped to a rectangle the bar no longer occupied, after the embedded browser changed its own display scale — something it can do when a monitor sleeps, a display is reconfigured, or the machine is locked and unlocked. The clip is now measured against what the bar actually paints, so a scale change makes it render slightly smaller rather than vanish. See ADR 0010.
- Changing any Flow Bar preference before ever dragging the bar silently reset *all* of them — the docked edge, snooze, "always show", compact mode — the next time OpenVoice started.
Changed
- The fold and unfurl are animated rather than instant: the bar unrolls along its width and then fills out, and folds flat before sliding away. Honours
prefers-reduced-motion.
You can choose a speech model again, and one of the choices speaks 25 languages.
Added
- The Speech model screen is back, with three options. Standard is included with OpenVoice, is what you get by default, and works offline from the moment you install — nothing here changes for you if you are happy with it.
- Multilingual (465 MB, optional). 25 languages, detected as you speak, with nothing to configure. Same speed and size as Standard. This restores the non-English dictation that 0.5.0 removed.
- Light (112 MB, optional). English, and a sixth of the disk and memory of Standard — for machines that cannot spare the room. It is noticeably less accurate, and the screen says so.
Changed
- Downloads are opt-in and verified. Nothing is fetched unless you press Download. Every archive is checked against a fingerprint built into the app before it is unpacked, and an interrupted or corrupted download leaves nothing behind rather than a half-installed model.
- The privacy note names the exception again. 0.6.0 uses the network for exactly two things, both of which you start: a model you choose to download, and the update check you can switch off. 0.5.0 could say dictation had no network path at all; with a Download button that is no longer true, and the copy says so rather than quietly staying optimistic.
- Upgrading from 0.4.x fixes itself. Settings naming a Whisper model that no longer exists are migrated to Standard on load, instead of leaving the screen showing nothing selected.
Fixed
- Choosing a model you have not downloaded can no longer leave OpenVoice unable to transcribe. Standard is always on disk and cannot be deleted, and the app falls back to it if anything else is missing.
Dictation is faster and much more accurate, and OpenVoice now works the moment it finishes installing — no download, no waiting, no model to choose.
Breaking: OpenVoice transcribes English only in this release, and the model picker and language setting are gone. If you dictate in another language, stay on 0.4.4.
Changed
- A new speech engine: NVIDIA Parakeet TDT 0.6B v2, replacing Whisper. On the same 25 test clips, it transcribes in about half the time of the old "Fastest" model while making roughly a seventh as many mistakes — it is simultaneously quicker than the quickest option you had and more accurate than the most accurate one. There is no longer a speed-versus-accuracy choice to make, which is why there is no longer a screen asking you to make one.
- The model ships inside the installer. Previously a first launch downloaded weights before you could say anything. Now it works offline from the moment setup finishes. The installer is correspondingly larger — about 440 MB against 68 MB — but you download those bytes once, at a moment you already expect to wait, instead of on first launch when you are trying to use the app.
- OpenVoice no longer needs Python. The speech engine is now part of the application itself rather than a separate bundled Python process. One process instead of two, and nothing left running if the app is killed.
- "Sends nothing anywhere" is now literally true. That note used to carry an exception — "except to download a speech model you ask for". With the model in the installer, dictation has no network path at all.
- Silence stays silent. Whisper would occasionally invent words out of room tone; a two-second near-silent recording once produced the word "Jordan". The new engine returns nothing for silence, room tone and hiss.
Removed
- Other languages. Parakeet v2 is English-only, so the language setting is gone rather than left as a control that changes nothing.
- The Speech model screen, and choosing, downloading or deleting models. There is one model, it arrives with the app, and Settings states which it is.
- The confidence score on history rows. The new engine does not produce one.
Known regression
- Proper nouns that sound like ordinary words may come out as the ordinary word. Marking a dictionary term as a proper noun used to tell the decoder that, say, "Claude" was a candidate while it could still hear the difference from "cloud" — something no amount of after-the-fact correction can recover. The replacement mechanism exists in the new engine but needs a file the model release does not currently ship, so it is a follow-up. Everything else your dictionary does is unchanged.
Changing your dictation shortcut now works the moment you change it, and the settings that genuinely cannot be applied to a running app finally say so instead of leaving you to guess.
Fixed
- Changing the shortcut takes effect immediately, instead of doing nothing until the next launch. Picking a new key wrote it to
settings.toml and stopped there. The keyboard hook went on matching whichever key it had been handed at startup, so the new key did nothing at all and the *old* one kept working — while the Settings screen showed the new one as if it were already bound. The machinery to rebind a live hook had been written and was simply never called from anywhere. It is now called on save, and the Hub's instructions and the Flow Bar's idle pill update to name the new key at the same moment, so nothing on screen can point at a key that no longer works. - "Hold to talk" and "Press to start and stop" switch without a restart too. The activation style was read once, when the session loop started, so changing it wrote a setting that nothing went back to read. Both settings used to carry the words "Takes effect after a restart" as a hint; neither offered a way to restart, and neither warned you when you closed the window.
Added
- Settings tells you when a change really does need a restart, and offers to do it. The speech model, the microphone, the maximum recording length and whether recordings are kept are all decided when the engine starts, and until now changing one from the Settings screen was silent — it was saved, it was not in force, and nothing said so. A banner now names exactly what is waiting and carries a Restart now button. It is computed by comparing what the running engine actually started with against what is on disk, so a setting that reloads in place — your shortcut, language, dictionary and profiles — can never appear there and train you to restart for no reason.
The release that actually reaches people. 0.4.1 and 0.4.2 were built and never published, so every download since 27 August served 0.4.0 — this fixes the distribution path that caused that, and the first-run failures that made a fresh install unusable when its model download did not complete.
Fixed
- The download button on the README now always serves the current build. It was a hand-written URL naming
v0.4.2 — a version that had never been tagged and never been built, so the link 404'd. The site's button, separately, named v0.4.0. Anyone downloading OpenVoice got a build from before two releases' worth of Flow Bar fixes, or nothing at all. Every release now also publishes the installer to a fixed download tag under a fixed filename, and documentation links there; CI fails if a version-pinned installer URL reappears in the docs. - Releases are published rather than left as drafts.
release.yml marked every release a draft, to be un-drafted by hand once the installer had been checked. Two releases were built and neither was ever un-drafted, so neither reached a single user, while the updater went on offering a months-old build. A release nobody can download is the same thing as a release that was never made. - A failed first-run model download is no longer permanent. The engine fetched its weights at startup and refused to start without them, and the Models screen's Download button required a running engine — so a first run interrupted by a dropped connection or a rate-limited request left the app answering "The speech engine is still starting. Try again in a moment." from the one screen that could have fixed it, on every subsequent launch, forever. Downloading no longer needs the engine, the startup error offers Try again, and a model fetched by hand starts the engine when it lands.
- Download progress is visible again. Every progress message was written from inside the
guard_stdout block that keeps a library's stray print out of the protocol stream — so the host received none of them, and the ticks were dumped to stderr afterwards as "suppressed stdout during load". A first run showed "Starting the speech engine" for the whole of a ~150 MB transfer with nothing moving. The Models screen shows a percentage on the button too, which it previously could not: progress was reported through the same channel as engine readiness, and readiness outranked it. - A dropped connection during a model download is retried. Three attempts with exponential backoff. The host deliberately does not retry a request that already reported progress, so nothing else in the system would have — and the Hub now explains a network failure in those terms rather than pointing at the log file.
- Advertised model sizes were about half the truth.
base.en was listed at 75 MB and transfers 148; small.en at 250 MB and transfers 486. Measured against the live repositories.
Changed
- The frozen sidecar now bundles `hf_xet`. huggingface_hub imports it from inside the function that decides whether to use it, so PyInstaller's static analysis never saw it and every installed copy fell back to plain HTTP downloads — slower, and far less able to resume, on exactly the transfer a fresh install cannot afford to lose.
This release completely overhauls the design system typography with bundled modern font packages (Geist Sans, Geist Mono, Inter, JetBrains Mono), introduces GPU-accelerated harmonic kinetic wave loading animations and ticking ellipsis, adds live UI reactivity (smooth WPM count-ups, auto-refreshing timestamps, real-time LED microphone VU meter), eliminates the Flow Bar rectangular window clipping artifact, and removes clutter from the idle Flow Bar.
Fixed
- Flow Bar translucent rectangular border artifact eliminated. Removed the outer drop shadow on
.flowbar that was bleeding outside the 40px rounded pill into the zero-margin Win32 GDI clipping region, strictly constraining shadows to an inner specular highlight (box-shadow: inset 0 1px 0 0 rgb(255 255 255 / 8%)) and ensuring zero pixel bleed beyond the pill curvature. - Streamlined idle Flow Bar visual design. Removed the redundant
· Prose / · Code profile tag in the idle Flow Bar state to maintain a clean, distraction-free indicator displaying exclusively Hold [Right Ctrl] alongside the microphone trigger.
Added
- Modern bundled typography system. Integrated
@fontsource-variable/geist, @fontsource-variable/geist-mono, @fontsource-variable/inter, and @fontsource/jetbrains-mono locally into the app bundle with full OpenType feature support (tnum, calt, kern, liga, cv02-cv11), replacing system defaults with crisp, high-legibility typefaces. - Kinetic harmonic loading dots and ticking ellipsis. Implemented GPU-accelerated harmonic 3-dot wave bounce (
<LoadingDots />) and synchronized kinetic ticking text (<TickingEllipsis />) during speech engine initialization, model downloads, and audio transcription. - Live audio reactivity and zero-render LED VU meter. Added a real-time 12-segment green/amber/red LED VU meter in Settings with direct microphone testing mode (
<MicTestMeter />), smooth numeric interpolation (useCountUp), and self-updating relative timestamps (useLiveTimeAgo).
This release fixes the Flow Bar context menu clipping into the taskbar or screen edge when the bar is positioned near the bottom of the display, introducing upward menu placement with zero-pixel pill displacement and precise native window region masking.
Fixed
- The Flow Bar right-click menu flips above the bar when parked near the bottom of the screen. When the bar was parked near the bottom edge of the display or just above the Windows taskbar, opening the 8-item context menu caused it to hang downwards past the display boundary and behind the taskbar, clipping options and swallowing clicks. The overlay now dynamically inspects available vertical screen headroom: if the bar is in the lower half of the screen (
top >= 340px), the menu opens upwards above the pill so all actions remain visible and accessible. - Desktop dead zones and menu clipping eliminated with symmetrical 640px window canvas and region masking. The transparent overlay window canvas is expanded to a symmetrical 404×640 canvas with the pill centered at
PILL_TOP = 300px (providing 300px headroom above and below). Win32 interactive region masking (SetWindowRgn) and overlay_set_shape dynamically track the active bounding box whether the menu opens upwards or downwards, ensuring the larger canvas never creates invisible click-blocking dead zones over background applications. - Smooth entrance micro-animations and zero-pixel pill displacement. Opening or closing the menu in either direction keeps the pill firmly anchored with zero pixel jitter or layout shifting. Menus enter with a snappy 140ms ease-out scale-and-fade micro-animation anchored to the respective top or bottom edge of the pill.
This release closes out the Flow Bar's geometry problems and its z-order problem in the same pass, then rebuilds it as a combination of Wispr Flow's and superwhisper's floating bars: a compact indicator, edge docking, and a menu that is a control surface rather than two links to the Hub.
Added
- The Flow Bar can be shrunk instead of hidden. "Compact bar" in its right-click menu reduces it to a small indicator that still shows the status light, and the waveform while you are speaking. Hover it and the full bar comes back. The choice is remembered between restarts. Hiding a bar you cannot see is the one thing that stops it telling you your microphone is open; making it smaller does not.
- Drag the bar to the left or right edge of the screen and it stands up. It reorients into a vertical column so it takes a strip of the margin rather than a bite out of whatever is maximised behind it. If something goes wrong while it is docked it unfurls back to horizontal to show you the message, because a sentence you have to tilt your head to read is not one you will read.
- Click the bar to start dictating. The shortcut is still the fast way; this is the discoverable one. Clicking to dismiss the menu does not start a recording, and a click that turns into a drag is a drag.
- The Flow Bar's menu is now a real menu. Start or stop dictating, paste the last transcript, jump to your transcript history, microphone or settings, switch between the compact and full bar, hide for an hour, or only show it while dictating. Each destination opens the screen it names.
- The tray icon has a Flow Bar submenu: Show Flow Bar, Hide until I dictate, and Reset Flow Bar position. Every control for the bar used to live in a menu on the bar itself, which is a dead end the moment the bar is not somewhere you can reach it — snoozed, switched off, dragged into a corner, or behind another window. The tray cannot be hidden.
Fixed
- The Flow Bar no longer disappears behind other windows after a few minutes. Nothing was hiding it. Windows silently strips the always-on-top flag from other windows when an application takes the screen fullscreen — a video, a game, a screen share — and never puts it back. No event is raised, and the bar still reports itself as visible, so it sat at the bottom of the stacking order with everything painted over it until the app was restarted. It now notices within two seconds and puts itself back, and records each time it had to, so the next person to report this has a number rather than a feeling.
- The Flow Bar no longer claims to be ready while the speech engine is still starting. During the first-run model download, the several seconds of loading after it, and permanently after the engine failed to start, the bar displayed "Hold Right Ctrl" — an invitation to press a key that was going to do nothing. It now shows the download's progress, says when it is starting, and says so in red when the engine is not going to come up.
- The Flow Bar no longer jumps up and to the left when you press the hotkey. The bar was held still while it changed size by moving its window in the opposite direction and letting the pill re-centre itself inside it. That only works if the move and the re-centring are drawn in the same frame, and they are produced by two processes on two schedules — so for the gap between them the pill was drawn at its old size inside the window's new position, displaced by exactly half the size change. Idle to listening, that is 67px left and 22px up.
- Dictated sentences now capitalize "I" and the days and months. Whisper transcribes the pronoun and every weekday and month name in lowercase, and nothing downstream lifted them, so "so i think we should ship on friday" landed exactly like that in your document. Sentence-initial words were already handled; these are the words English capitalizes wherever they fall. Names of people, places and products are deliberately still left alone — guessing those needs knowledge the formatter does not have, and a wrong guess is worse than the lowercase it replaced.
- "fifty percent" is no longer written as "fifty%". The spoken word
percent was substituted for % wherever it appeared, which is right after a number and nonsense after a word. It now converts only when a digit precedes it, so "up 50 percent" gives you up 50% and "up fifty percent" is left as spoken. - The Flow Bar no longer flashes a cropped black rectangle when you press the hotkey. For a moment as the bar switched into listening, it could appear as a square black box with the pill clipped inside it — one rounded corner showing, the green edge running into a hard cut, the timer sliced through. Two separate races caused it, and both are closed: the pill was being painted at its new width before the window had been resized to hold it, and the window was being resized before the surface it is drawn on had caught up. The pill's size is now read off the window rather than told to it, so the two cannot disagree.
- The Flow Bar no longer drifts after you drag it near a screen edge. The animation that slides the bar the last few pixels into a snap was being denied permission to move the window, silently, on every frame — while the bar went on recording the position it believed it had reached. The next time it resized, it measured from somewhere it had never been.
- The bar is now sized correctly at startup with a non-default shortcut. The first resize after launch was dropped, leaving the window at the size declared in configuration rather than the size the pill actually needed.
- The bar no longer jumps up and to the left when you press the hotkey repeatedly. Once you had clicked the bar even once, every automatic resize after that was mistaken for you dragging it, and the bar's remembered position was rebuilt from a size it had not reached yet. It moved about 57px left and 22px up on release and back again on the next press.
- The bar no longer stays stretched after you cancel a dictation. Cancelling while the bar was still growing into its listening size left the window at that larger size with the idle content inside it.
- A click on the bar that does not move it no longer counts as a drag. The next time the bar was placed automatically, that placement was saved as a position you had chosen.
- The bar no longer slides up and to the left when you use the hotkey. The bar can tell its own movements apart from yours now, instead of guessing. When it guessed wrong, the ordinary resize that follows every press was treated as you having dragged the bar: it was snapped to a screen edge measured for the wrong size, saved to disk as your chosen position, and then slid there over a fifth of a second. It also left the bar's idea of where it belongs permanently wrong, so the next press started from the wrong place.
- The slide that follows a drag can now be interrupted. It ran to completion no matter what else happened, writing a position sixty times a second against everything else that moves the bar — so a dictation started mid-slide fought it, and the slide won.
- The Flow Bar can no longer steal focus from what you are typing into. The flag that prevents it was applied once at startup and then quietly cleared by the window toolkit, so on a running copy the bar was able to take focus — and losing the caret is how a dictation lands in the wrong place. It is now re-applied whenever the bar is shown.
This release is about the Flow Bar — the small pill that floats while you dictate, and the only part of OpenVoice most people ever look at. It can now tell you when something went wrong, let you throw away a dictation you did not mean to start, and it moves like something that was designed rather than resized.
Added
- Discard a dictation in progress. Press Escape, or use the new × on the bar. Previously, releasing the key always transcribed and injected: an accidental trigger had to land in your document before you could undo it. Escape has quietly worked for a while; nothing ever said so.
- The bar tells you when a dictation did not land. A failure and a clipboard fallback each get their own state on the bar, with the reason. Before this, both looked exactly like nothing having happened — you found out when you went looking for text that was not there.
- A confirmation when your words arrive. A single ring leaves the status dot as the text lands, so you can tell success from silence without turning your head.
- The bar glows while the microphone is open, and the glow tracks your voice. It is meant to answer "is it hearing me?" from the corner of your eye.
- More vocabulary out of the box: Vercel, Tauri, GitHub, pnpm, Claude, Anthropic, MCP, SDK, UX, CSS, HTML and a few more now come out spelled the way you meant. Only spoken forms that are not also ordinary English are claimed — a dictionary that confidently rewrites real words is worse than none.
Changed
- The bar moves properly. Releasing a drag near a screen edge now slides into place instead of jumping, and the bar shows you when it is about to snap. Its width changes ease between sizes rather than snapping between them.
- The waveform behaves like a meter, rising fast and falling slowly with a held peak, instead of rising and falling at one flat rate. It also uses fewer, clearer bars — at the size the bar is actually seen, thirty-two of them read as a green smear rather than as your voice.
Fixed
- Dictating twice in quick succession no longer loses the second utterance. Releasing the key left the recording slot occupied for a few hundredths of a second while the audio was handed back. A press inside that gap was discarded outright — no recording, no error, nothing on screen — so you could speak a whole sentence into nothing and only discover it afterwards. The press is now held and honoured the moment the slot frees.
- Three dictations in quick succession no longer produce the wrong text. Only one utterance's audio was kept at a time, while the queue behind it could be several deep. The third recording overwrote the second's audio, so the second dictation was transcribed from the third one's sound — inserting words you said at a different moment — and the third then failed outright. Each recording now keeps its own audio until its turn.
- A dictation that fails now says so. Transcription and microphone failures reached the Flow Bar in the same instant it was told to go back to idle, so the failure was never drawn. Failing silently is the one thing a tool that types for you must not do.
- "I mean", "kind of", "sort of", "actually" and "literally" are no longer deleted from what you said. Filler removal treated these as noise, and the profile used for browsers, chat and mail removes them by default — so in most applications they silently vanished. They are not noise: "I mean it" became "it" and "kind of blue" became "blue", which is the opposite of what was said. Only tone words are removed now; anything that can change meaning stays. "you know", "like" and "basically" are still removed at the aggressive setting.
- "Literally" is no longer swallowed before the next word. It is the escape word that stops a voice command being interpreted — "literally new line" types the words instead of breaking the line — but it was consumed before *anything*, so "it is literally true" arrived as "it is true". It is now only treated as an escape when a real command follows it.
- A custom shortcut no longer breaks the idle bar. The bar was sized for exactly one shortcut — the default. Anything longer collided with the word "Hold" and was cut off. It is now measured from whatever your shortcut is.
- Error messages are no longer truncated mid-sentence. The bar was fixed at a width that cut most messages off after about thirty characters, which is reliably before the part telling you what to do about it.
OpenVoice can now tell you when a new version exists, manage its own speech models, and learn a word it got wrong. Several settings that had never done anything now do what they say — and one bug meant no fresh install could download a model at all.
Added
- Update checks. OpenVoice looks once per launch for a newer release and offers it; you press a button to install. It never installs on its own. Every update is verified against a signing key built into the app, so a tampered or substituted download is discarded without running — worth having while the installer itself is still unsigned. Switch it off in Settings → Updates and no request is made at all. This is the one request OpenVoice makes that you did not individually ask for, and the reasoning is in ADR 0005.
- Secrets are kept out of your history. API keys and tokens matching
privacy.redact_patterns are replaced with [redacted] before a transcript is saved or logged. Defaults cover OpenAI, GitHub and AWS credentials. The text delivered to your application is never altered — only the stored copy. - Press once to start, press again to stop. Holding a key is fine for a sentence and unpleasant for a paragraph, so there is now a choice under Settings → Dictation. Hold-to-talk is still the default, because it is the only mode where the microphone cannot be left open by accident. A session you forget to stop still ends at the maximum recording length.
- The dictation key can be changed. Eighteen keys to choose from, including F13–F24 for anyone with a macro key or a remapper. The Hub and the tray now name the key you actually bound instead of always saying "Right Ctrl".
- Teach it from a dictation that came out wrong. Every entry in Recent now has a Fix a word button. It shows what OpenVoice actually *heard* — which you have never been able to see before — and you click the words it got wrong, type what you meant, and save. It applies to the very next thing you dictate. Knowing that
kubectl arrived as "cube control" is most of the work; guessing it from the tidied-up output is not. - Models can be downloaded and removed from the app. The Speech model screen now shows which models are on your disk, how much room each takes, and the total. Download fetches a model without switching to it or restarting, so you can get the weights first and decide later; Delete gives the space back. Deleting the model in use is refused rather than left to fail later.
- Recordings clean themselves up. If you turn "Keep recordings" on, they are now deleted after a week by default, adjustable to a day, a month, or never. Audio is roughly 2 MB a minute and previously nothing ever removed it. Your history is separate and unaffected — turning recordings off, or having them swept, never touches your transcripts.
Fixed
- Downloading a model works at all. No model could ever be fetched by an installed copy — including the very first download on a fresh machine, which made a new install unusable unless it happened to find weights already cached. Offline mode is read once by the download library when it starts and could not be lifted afterwards, so every fetch failed claiming outgoing traffic was disabled. This was present before this release and was invisible on any machine that already had the models.
- The Speech model screen looks in the right place. It reported "using nothing on this computer" and offered to download models already present, because it always looked under
%APPDATA%\OpenVoice\models. That is only where weights live for a standard install; if OPENVOICE_PYTHON or HF_HOME is set the app uses a different cache, and the screen now asks the running engine which one it is actually using. - A model you have chosen but not downloaded no longer claims to be "In use". It reads "Selected — not downloaded", next to the Download button that will fetch it.
- A failed update check reads like information rather than a fault. Nothing is wrong if no release has been published yet, or if you are offline, and the message now says so while keeping the underlying detail.
- "Maximum recording" now has an effect. Choosing 5 minutes still cut the recording off at 2: the whole
[limits] section was written to disk and then ignored, because the engine started the session machine with built-in defaults instead of your settings. - "Keep recordings" now keeps recordings. The toggle had nothing behind it. Turned on, captured audio is kept under
%APPDATA%\OpenVoice\audio from the next restart instead of being deleted after each decode. It remains off by default, and it is still the only way audio ever persists. - The Speech model screen lists what the app can actually load. The model list was written down twice — once in the speech engine, once in the interface — and the two could disagree about which models existed and how large they were.
Removed
model under [config] in settings.toml, which was read by nothing. The model has always been the top-level model key, and that is unchanged. Your existing settings file loads as-is; the stale key is ignored, not deleted.
A fix release. The visible complaint was that the Flow Bar could multiply into several copies of itself, each drawn inside an opaque rectangle. The bar was innocent: OpenVoice could be running several times over, and each copy was showing its own.
Worth upgrading to even if you never saw duplicate bars — the same fault meant several copies could be holding the microphone, decoding the same audio, and writing the same history database at once, none of which is visible from the interface.
Fixed
- Launching OpenVoice while it is already running no longer starts a second copy of it. Closing the Hub hides it rather than quitting, which is deliberate — a dictation tool that stops working when you close its window is not available when you need it. But nothing stopped the *next* launch from starting a complete second app: another global keyboard hook on the same chord, another Flow Bar, another 1.6 GB of weights, and two processes writing the same history database and overlay placement. Every hotkey hold then opened the microphone once per copy, transcribed the same audio that many times, and raced to inject the results into each other. Launching again now raises the window of the copy already running, which is what it was always meant to mean. The duplicates were easiest to notice after "Hide for an hour", because that kept each new copy's bar hidden until a dictation forced it on screen — so they arrived all at once, on one keypress, looking like one window gone wrong.
- The Flow Bar no longer flashes a dark rectangle around itself. The overlay window is transparent, but transparency was applied from a script that runs after the first paint, while the app's black canvas was painted the moment the stylesheet loaded. The window filled itself in for the whole gap between the two. The canvas is now something each window opts into, so the overlay has nothing to fall back to; the window is also declared at the size of the idle pill rather than 280×52, so it no longer appears oversized and then shrinks.
- The Flow Bar's right-click menu now closes when you start dictating. Its only dismissal path was losing window focus — which cannot happen to a window built never to take focus. Opening the menu and then dictating left the panel standing open behind the bar as an opaque block, with no way to click away from it.
- Auto-placement is no longer recorded as a position you chose. The Windows move loop swallows the mouse-up that ends a drag, so a click on the bar that moved nothing could leave it looking permanently dragged; the next time the app placed the bar itself, that placement came back as a drag, got snapped to an edge, and was saved.
The first published build. Hotkey → capture → transcribe → format → inject → history works end to end on Windows, and it installs from a single .exe with no Python, no Rust and no Node on the target machine.
This is an alpha, and the version number is the promise: there is no upgrade path from anything, and nothing here is a compatibility commitment yet. Two things to know before installing, both covered in the README: the installer is not code-signed, so SmartScreen will warn about an unknown publisher (verify the published SHA-256 instead), and the bundled speech engine is CPU-only — using an NVIDIA GPU still means running from source.
Added
- Sound feedback. A short tone on starting a dictation, and a distinct one on a successful finish — both synthesized with the Web Audio API rather than shipped as audio assets, so there's nothing to license or bundle. Off/on in Settings (
Config.sound_enabled, default on). The finish chime is deliberately withheld on a clipboard-fallback completion, which settles to the same "idle" state a clean success does — playing it there would tell the user a dictation landed cleanly when it didn't. - A language setting.
ov-asr/the sidecar have accepted a forced-language parameter since day one, but it was hardcoded to English at every call site in ov-app and ov-cli with no way to change it. Now a real setting (Config.language), exposed as a dropdown in Settings, defaulting to English rather than auto-detect — Whisper's auto-detection looks at only the first ~30s of audio and is measurably less reliable than being told the language, which matters most for exactly the case this app is built around: a short, single push-to-talk utterance. Unlike the model, changing it needs no restart — it's a per-request decode parameter, not something baked into loaded weights, so it reloads the same way the dictionary does. ov-cli gained a matching --language flag (auto for auto-detect). - An installer.
OpenVoice_0.1.0_x64-setup.exe, 47 MB, with the speech engine inside it. No Python, no virtualenv, no Rust on the target machine. Built by release.yml on a tag, with a SHA-256 published beside it. - A frozen speech engine. PyInstaller packages the sidecar into a standalone binary (
scripts/build-sidecar.ps1). The build is not considered finished until the frozen binary answers a real request on its protocol — a missed hidden import is invisible to PyInstaller and fatal at run time. - First-run model download, with progress. The weights are fetched on first launch and reported as a real byte count rather than a spinner. Previously the app simply looked frozen for the several minutes this takes.
- Screenshots that cannot go stale.
scripts/screenshots.mjs captures the README images from the running UI over the DevTools protocol.
Changed
- An installed copy keeps its weights under
%APPDATA%\OpenVoice\models rather than in a shared Hugging Face cache, so uninstalling reclaims the space. - A debug build always prefers the Python sidecar in your checkout; a release build prefers the frozen one it shipped with. Edits to
sidecar/ are no longer silently shadowed by whatever was last frozen. - The speech engine ships CPU-only. The CUDA libraries are 1.9 GB against 240 MB for everything else and do nothing without an NVIDIA GPU.
- The installed default model is now `base.en` (~75 MB), not `large-v3-turbo` (~1.6 GB). Pairing the heaviest model with a CPU-only install was the wrong default: worst latency, biggest download. Upgrading is a Models-screen action, already built, that downloads on demand.
- PyAV is no longer in the frozen build, cutting it from 240 MB to 173 MB.
engine.py reads its own WAV files (it is the only producer of them, and always writes 16 kHz mono 16-bit PCM) instead of going through faster-whisper's decode_audio, which is the only thing that needed PyAV's bundled FFmpeg. Verified against the real fixture audio, not just unit tests.
Fixed
- A dictated accented character, CJK character, or emoji could crash the sidecar or silently lose the transcript. Python's stdio only defaults to UTF-8 inside an interactive console; redirected to a pipe — exactly how the Rust host launches this process — it fell back to the Windows ANSI code page (
cp1252 on a typical install). Confirmed by actually spawning a piped child process: a character cp1252 can't represent at all (CJK, emoji) crashed the sidecar outright, and one it *can* represent as a different single byte than UTF-8 uses — which includes ordinary accented Latin text ("café", "naïve") and the smart quotes/dashes Whisper's own punctuation restoration commonly adds — wrote bytes the Rust side rejected as invalid UTF-8. The automatic retry couldn't help, because it decodes the exact same audio and hits the exact same failure again. Fixed by forcing UTF-8 on stdin/stdout/stderr explicitly, first thing in main(), rather than trusting the launching environment. Verified with a real piped subprocess in both directions (fails without the fix, survives with it). - Four idempotency bugs in the formatting pipeline, found by fuzzing a battery of realistic phrases through every profile twice (formatting is required to be idempotent on its own output — history replay and any "preview this phrase" UI depend on it): the
literally escape hatch only protected one word after itself, so a two-word command like literally fat arrow still had its second word (arrow) reinterpreted as ->; a resolved identifier (userName) got re-capitalized into UserName if already-formatted text was re-parsed; Doc::parse used split_whitespace, which silently discarded the literal newlines that "new line"/"new paragraph" produce on a second pass; and force-lowercase profiles corrupted a surviving SCREAMING_SNAKE_CASE constant into mAX_SIZE instead of leaving it alone. - The Flow Bar overlay had permissions it never used and must never use. It shared the Hub's full capability grant, including
core:window:allow-set-focus — directly contradicting the overlay's one architectural invariant, that it can never take focus. Nothing currently misuses this, but any future or compromised dependency running in that webview could have called getCurrentWindow().setFocus() on itself and it would have been permitted. Split into its own capability file granting only what Overlay.tsx actually calls. - The clipboard-restore decision after a paste couldn't tell "succeeded" from "failed, and the caller is now using the clipboard as a safety net" — both end up with the same text sitting on the clipboard, and the old check only asked "is the clipboard still ours," which is true either way. Refactored into a single, unit-tested decision (
should_restore) that also checks whether the paste actually sent. - A failed clipboard-paste recorded the wrong text. When injection fell back to leaving text on the clipboard, history and the
Outcome::ClipboardFallback payload recorded the *raw* ASR transcript instead of the *formatted* text — the thing that was actually placed on the clipboard. History could disagree with what Ctrl+V actually pastes. - A muted/disconnected microphone was recorded as "transcription failed."
AudioFailed (capture never even reached the transcriber) was mapped to the same Outcome::AsrFailed as a genuine model failure, so a dead microphone and a broken sidecar were indistinguishable in history. Split into its own Outcome::CaptureFailed, threaded through the Rust↔TypeScript wire contract. - The engine could report "idle" while still transcribing a queued session. Holding the hotkey again while a previous utterance was still being processed queues a second session; if that second session ended as a fat-finger tap, silence, or a capture failure, the engine unconditionally announced
Idle — even though the first session was still actively in flight. Now only announced once the whole pipeline is genuinely empty. - A NaN or negative silence threshold silently disabled muted-mic detection. TOML happily parses
nan/-nan float literals, and rms < silence_rms is false for every possible rms when the threshold is NaN (or negative, since RMS is never negative) — completely and silently turning off the check with no error anywhere. Config validation now rejects both. - A stale, incorrect
tauri_build::build() assumption meant a fresh checkout (a new contributor, or CI, neither of which have ever run the sidecar freeze step) would fail the very first cargo test/clippy/check it ever ran — bundle.resources is validated against the filesystem unconditionally, not just for an actual tauri build. This had been invisible all session because the frozen sidecar already existed locally from manual testing. build.rs now creates an empty placeholder for ordinary dev builds, but still hard-fails a real release build if the frozen sidecar is genuinely missing. - A failed short dictation left nothing on the clipboard at all. Only the long-text/clipboard-paste path had a fallback that left text behind for
Ctrl+V on failure; the short-text path (send_unicode, used for the overwhelming majority of dictations, since most utterances are well under the paste threshold) returned the error directly with no fallback. A failed short dictation was silently gone. Both paths now leave text on the clipboard on failure. - The clipboard fallback for a failed paste was silently undoing itself. When injection can't type text into the target app directly, the text is left on the clipboard so
Ctrl+V still works — but a background thread was restoring the user's *previous* clipboard contents exactly 1 second later, regardless of whether the paste actually succeeded. That 1-second window existed to solve a different race (giving the target app time to read the clipboard) and was never long enough for a *person* to notice a failed paste and react to it. Worse, the restore fired even down the known-failure path, undoing the exact fallback the error branch had just set up. Fixed two ways: the hold is now 15 seconds, and the restore no longer runs at all when the paste chord itself failed to send. A failed injection also now shows a notice telling the user to press Ctrl+V — previously the only way to discover an injection had failed was to spot a non-"delivered" badge in History afterward. - Injection failures were completely invisible in the log.
Effect::Inject logged nothing on error — the only trace was the in-app notice and a history badge, neither of which helps once the moment has passed and someone is trying to reproduce an intermittent failure. It now logs the error directly. Also added a check for a real, previously undiagnosable class of bug this surfaced: transcription and formatting can take several seconds, so the window that was foreground when the user started speaking is not guaranteed to still be foreground at injection time. That gap was silent; it now logs a warning naming both the app the user spoke into and whatever is actually foreground at injection, so a report of "it sometimes doesn't paste into app X" is diagnosable from one log line instead of a guess. - History no longer shows a green "delivered" badge on every single row. It is the default, expected outcome for nearly every dictation, so repeating it down the whole list was decoration, not information — and diluted the one place green is supposed to mean something (the microphone is open) into noise. A badge now appears only when an outcome needs attention.
- The sidecar no longer flashes a console window on every start and restart.
beforeBuildCommand resolved its path against whichever directory invoked the build, so tauri build failed depending on how it was called. Build commands now live in the root package.json, which npm run finds by walking up.- The design-system sheet documented
--faint as #6b6b6b at 3.4:1 — a value the tokens had already moved away from, on the one page whose job is to be correct about them. - Excluding PyAV initially broke
import faster_whisper outright — faster_whisper/__init__.py imports it unconditionally, whether or not decode_audio is ever called. Caught by actually running the frozen binary's probe, not by reading the diff. Fixed with a stub module installed only when the real package is absent, scoped to the frozen entry point. ensure_model's request nested "model" under a literal "params" object, which the sidecar's own wire format treats as a param named params — the model name never actually reached the handler, which silently fell back to the sidecar's already-configured model instead. Harmless today only because ensure_model has never been asked for a different model than the one the sidecar was launched with. Found by writing a real end-to-end test against the frozen binary rather than trusting the unit tests already in place.