Every version, in plain English. You’ll find the same list in the app under About.
0.19.2current
The recognition server no longer mixes your words into other people's recordings.Your personal dictionary and replacements now apply only to your own dictation and to files you open yourself. Requests from other programs get exactly what the program sent in its prompt. Before, on noise or silence a word from your personal dictionary could end up in the answer.
You can say how many people are speaking.A program calling the server can pass the number of voices: "exactly N" or "at most N". Voices merge into one less often.
The request feed no longer repeats the same line.Every row used to carry an identical note about the dictionary.
On silence the server no longer answers with the prompt.If recognition returns only words from the calling program's prompt (it happens on noise), the answer comes back empty with a flag. The cost: a real reply that is just one word of the prompt disappears too.
0.19.1
Removed the needless padlock from the bar.The "where do my voice and text go" icon now shows only when audio or text really leaves this computer: to another machine or to the cloud. While everything stays with you, the bar carries nothing extra.
0.19.0
When the microphone is silent, the app tells you right away.After three seconds of silence the bar shows a "Can't hear you" plate with a "Check the microphone" button: you used to speak to the end and get an empty text.
Microphone order.In the sound settings you can build a list of backups: when the main one is not around, dictation takes the next one that is connected. A Bluetooth headset is not picked on its own while anything else is plugged in: its microphone switches the headphones into call mode and ruins the sound. If you want that one, put it in the list.
Where your voice and text go is visible on the bar.A small icon says whether the audio stays on this computer, goes to a server on your network or to the cloud, and the same for the text when AI processing is on. Hover it for the plain-words version.
"What went to the AI" in History.For a dictation the AI processed, you can see which model edited it and open the text from before. If you do not need originals kept, turn off "Keep the text from before AI processing": the ones already saved are erased from disk, the final texts stay.
Fixed a word by hand? The app offers to remember it.Correct a name, a term or an abbreviation in History and it asks whether to add it to the dictionary, so it is recognised right from then on. Ordinary words and numbers are not offered.
"Paste the key from the clipboard" on the Pro screen.Copied the key from an email — a button appears on the Pro screen, no typing needed.
The settings window and every label were reread.Grey and coloured text is readable in all three themes; keyboard focus is visible everywhere; units follow your language ("1.6 GB", "850 ms", "2 min 3 s"); the "Close" button no longer sits on top of dialogs; in a narrow window the side menu hides by itself; the model recommendation is worked out from your hardware and no longer mixes "download" with "best of the installed".
A model error is visible on the bar.A red dot with "!" instead of silence, and a click on the gear opens the model settings.
Hotkeys.One and the same set of choices in the wizard and in Settings; warnings show before you choose; the Ctrl+Win combination is added.
The "Licenses" window now opens with a plain explanation above the original texts,and "What's new" no longer carries working notes.
0.18.1
The app now says what Pro would do — about your text, not in general.When the passage you just dictated carries no full stop at all, or still has its "um"s and "basically"s, the hint names that. Once per launch, and only when the text really gives it something to say: on a short note or a clean sentence it stays quiet.
0.18.0
Dictate on one computer, recognise on another — with no setup.The recognition server can open a second, encrypted door and hands out a connection code: one line that already holds the address, the key and the node's certificate. On the other computer it goes into a single field. A tunnel (Tailscale) is no longer required on your own network — only for reaching the machine across the internet.
The network port opens only when you switch it on, and it is off by default: a node nobody invited outside still answers only programs on its own machine.
The access key can no longer be spent by accident.The app refuses to send a key to an unprotected address and says what to change: it used to go out in the clear, the node declared it compromised, and all the person saw was that recognition stopped working.
The opening words of a dictation survive a long pause.A key latched down in the background by another program was costing two seconds off the first dictation after every break.
The hint under the server address no longer suggests an unprotected one.
0.17.0
The recognition server has stopped misreporting itself. While it transcribes somebody else's recording, your window no longer runs a "Transcribing…" bar: that job is not your dictation and must not look like one.
If the listener dies after it came up, the panel notices and names the reason. It used to show a green "running" and an address nothing answered on.
When the Corp licence runs out, the server actually shuts down. It used to promise that and not do it, so a connected program kept knocking for ever.
A damaged recording from somebody else's program is no longer blamed on your node: it is told "this file cannot be read" rather than "come back later". The old answer took the node out of rotation for a whole cooldown.
Your Windows account name no longer travels into other people's logs. The error text was handed back as-is, and a temp-file path sat inside it.
"Ready to take requests" now means what it says. The most useful state is reported separately: ready, but the first request waits for a model load — tens of seconds, and worth knowing before you ask.
A refused key — the commonest "it does not work for us" — now reaches both the on-screen feed and the log. It used to leave no trace at all, so you learned about a broken integration when a colleague phoned.
The request feed survives a restart: it is rebuilt from the log that was already being kept. No recognised text comes back with it — none was ever there.
There is a "Check the node" button. It sends a real request through your own address and key and covers the whole path at once: port, key, queue, engine.
Beside it, today's tally: answers, refusals, failures, and how long the node usually takes. "Nobody has ever connected" and "340 requests this morning, quiet now" no longer look the same.
The tray tooltip says whether the node is listening or down, without opening a window.
Your brand glossary no longer disappears when a connected program sends its own: the two lists are merged. Yours used to be switched off entirely, and silently.
The node says in its answer when the engine could not be given the language that was asked for, and when the text passed through your dictionary and replacements.
A new "keep the model ready" switch: the first request stops waiting for a load. Off by default — a resident model holds video memory the whole time the server runs.
The longest recording a caller may send is a setting now, thirty minutes by default instead of a hard-wired ten.
An access key can be issued with an end date and an hourly request limit. An issued key used to work for ever, and so did a leaked one.
A key cannot be issued without a name. A month later, "unnamed" in a list of three is a key nobody dares revoke.
Everything a second program needs — address, path, length ceiling and request shape — sits in one block with a "copy all" button.
Without a Corp licence the section is now the switch, the explanation and a link. It used to show the whole console, every control of it dead.
The section is findable by settings search at last — neither its settings nor even its own name used to come up.
The "show the recognised text" switch goes back to off on every start. That is other people's speech, and a switch left on went on showing it.
A request with no key no longer pushes tens of megabytes through the machine before being turned away.
0.16.11
When an update check fails, the app says so. The About screen used to read "you have the latest version" instead: the time of the last successful check is restored at startup and outlived every failure after it, so an updater that could not reach the server looked perfectly healthy.
0.16.10
The "Going through the server" panel no longer claims the machine is quiet while it is busy with your own work. It names what it is doing: your own dictation, your own file, a speed test, or warming the model up. The node used to look idle and refuse other people's requests at the same time.
The Copy button on an issued access key can no longer come back empty: the app now confirms the key really reached the Windows credential store before promising it can be copied.
When one key of your hotkey combination is stuck down — some keyboard drivers and macro tools do that — the log now names it. Until now every dictation quietly waited an extra two seconds with nothing to explain it.
0.16.9
An issued recognition-server key can be copied again — the row used to offer nothing but deletion. The key lives in the Windows credential store rather than in a settings file; keys issued by earlier versions have no copy kept, and the row says so.
Errors coming back from cloud recognition services are now readable in your own language. A dropped network and a microphone held by another app no longer look like a fault in the app — it says what actually happened.
A call to a cloud service can no longer run longer than its allowance: the deadline now covers the whole call rather than each attempt separately.
The model comes back more reliably: an auto-unload tick that outlived the watch no longer takes it away, nor overwrites what its replacement had just restored.
Recognition-server node: two recordings arriving in the same second count as two, not one.
Earlier versions (52)
0.16.8
The model comes back on its own the moment an auto-unload program closes. The app never noticed those programs closing: the screen kept saying "paused while … is running", the model stayed out of memory, and a dictation had to be started by hand to get it back.
The same cause made auto-unload fire only once per session — a second run of the same program was treated as already known and the model stayed loaded. It now works every time.
Turning the auto-unload switch off brings the model back immediately, even while the program is still open.
On the Recognition server screen the device is no longer reported as "cpu" before the engine has been loaded: it says the device is not known yet, and the slow-performance warning no longer appears without cause.
0.16.7
The Recognition screen now says WHY the model is not loaded: unloaded after idling (and after how many minutes), paused while an auto-unload program is running (naming it), unloaded by hand, failed to load (with the error), or not downloaded yet. One grey line used to cover all of them, so there was no way to tell whether anything needed doing.
The same reason appears in the tooltip of the "Model is not loaded" pill in the settings header.
Internal libraries updated.
0.16.6
The update notes are now written in the app's language and list what actually changed instead of pointing at another window.
A link to the full list of changes for that version sits under the notes.
0.16.5
"Reset all settings" now checks that the cloud-service keys were really removed and says so when the OS credential store kept them.
The repeat, cancel and show-window hotkeys: when the setting fails to save, the system is no longer left listening for a combination the screen does not show.
0.16.4
Junk-phrase cleanup no longer cuts neighbouring words when the text contains letters such as "İ".
A model download whose size the server did not report now reads as a download, and a second click no longer starts it again.
The copy buttons on the recognition-server card say so when the clipboard refuses.
The network switch no longer blinks "off" while the server restarts.
The focus ring on the selected theme is visible from the keyboard.
0.16.3
A recording in another language can no longer be recognised as Russian — that used to produce a translation instead of a transcript.
Speaker separation now works with the cloud engines that return no timings, and the recognition server no longer answers with an empty text.
When the key field holds more than one key, the app asks which one to use instead of picking one itself.
A model that was just warmed up is no longer unloaded straight away.
0.16.2
A recording that opens with an ordinary greeting no longer loses its first sentences.
Stray phrases like "Laughter" no longer stay in the speaker-labelled text and the subtitles when a similar real word appears nearby.
Server-mode settings are applied in the order they were changed.
0.16.1
The subtitle export button now appears on every file transcript, not only on the ones with separated speakers.
The names you give the speakers now reach the saved subtitles too.
The duration shown in the history matches the length of the source file.
0.16.0
Subtitles from a transcript.Any history entry can be saved with timecodes — SRT, VTT or timed text. Lines carry speaker labels, and speakers can be renamed to real names that reach both the document and the subtitles.
Speaker separation is more accurate when you know the number.The "Voices" card lets you say how many people are on the recording; by default we work it out.
Long recordings transcribe fully on every engine.Parakeet no longer hits its length limit, and GigaAM's piece seams no longer leave a repeated word.
Less junk in the transcript.Loop cleanup now covers every engine, not just whisper, and reaches the speaker-labelled text too.
The trial shows what you are paying for.One switch in the setup wizard turns on AI text processing for the 14 days: the model downloads in the background and switches itself off when the trial ends.
Small things that got in the way daily.The insertion method is honoured for manual inserts too; edits in the result window are saved to the history; a single file can be stopped; file transcripts survive the retention window and are found by file name.
0.15.8
The record day in statistics is recomputed from dictation.If file transcriptions raised the record in earlier versions, it corrects itself on the first start.
The "Voices" model set no longer offers a 31 MB download to users who cannot switch it on.
The recognition server restarts reliably when its settings change: the port is no longer held by its own outgoing listener for a fraction of a second.
0.15.7
Speaker separation is now a Pro feature.File transcripts with "Speaker 1: …" lines no longer need the Corp plan: the "Voices" model set and the switch are open to every Pro user.
GigaAM handles long recordings.A file longer than half a minute could stop with an error; it is now recognised in pieces cut at quiet points, on one shared timeline.
Statistics count dictation only."Words today", the daily chart, the model breakdown and the record day are no longer inflated by file transcriptions — the files stay in the history.
The recognition server now starts and stops right away on settings import and "reset all settings", not after a restart.
0.15.6
The junk-phrase filter now covers speaker-labelled text too.Lines like "Subtitles by…", which whisper sometimes invents over silence, were cut from plain text but survived in the "Speaker N: …" text and in the history. Not any more.
0.15.5
File transcriptions are now in the history.A transcribed file used to live only in the result window — close it and it was gone. It now sits in History & Stats like any other entry, with the file's name on the card: searchable, copyable, starrable.
0.15.4
Long recordings get their punctuation back.On files longer than 30 seconds whisper could return the whole text without a single period or comma — commas, periods and question marks are back, the words unchanged. Russian and English.
0.15.3
Files with speaker separation come back faster.Speakers are now worked out at the same time as the text instead of after it: a five-minute recording takes 29 seconds instead of 49. Same result, just sooner.
0.15.2
One button on the recording bar instead of two identical ones.A "stop" square on the left sat next to an identical "cancel" square on the right and the two got mixed up. The stop button is gone: a click on the bar itself or the hotkey stops a recording. The cancel button stays and wears a trash icon now, so it is obvious it throws the take away.
0.15.1
Speaker separation now works with GigaAM too.With that engine a file transcript used to come out as one block without "Speaker N" labels — the labels are there now.
The "Transcribing file…" tooltip no longer covers the timerabove the bar.
0.15.0
Speaker separation.A new "Voices" model set (31 MB, Settings → Recognition) labels who spoke when: file transcripts come out as "Speaker 1: …", "Speaker 2: …" lines, and the recognition server returns speakers on request (the diarize field). Works with every local engine. Best on clean recordings of up to 4–5 people. A Corp plan feature.
The Corp planlists both of its features — the recognition server and speaker separation — first.
0.14.4
A new key goes in over the old one.An active licence now has an "Enter another key" button — for example to move from Pro to Corp. Nothing has to be removed first.
The key can be pasted as a whole line from the letter.Extra text around the key no longer blocks activation.
The app shows how to get Corp.The plan section and "Recognition server" link to the team prices and enquiry form.
Corp leads with its main feature.After activating a business key, "Recognition server" is the first card — in the welcome and in the feature list.
0.14.3
The key field explains a refusal.When a key is not accepted, the field now says why instead of staying silent.
"Recognition server" follows the licence immediately.The toggle and the server state change the moment a business key is activated or removed, no restart needed.
0.14.2
A business licence is called by its name.After entering a Corp key the "Pro" screen said "Pro", and some features stayed locked until you reopened the section. The plan now reads "Corp", the licence card appears right after activation, and every Pro feature is open from the first second.
The key field explains itself.The hint says a Pro key and a Corp key go into the same field, and a "Where do I get a key?" link sits next to it.
0.14.1
Text arrives about 200 ms sooner, on every dictation.After you released the hotkey the app was recording the tail twice: once while it made sure the key really had come up, then again for good measure. That tail is now counted once. Trailing syllables are still caught.
Russian speech no longer arrives in English.On short phrases the recogniser sometimes got the language wrong, and then returned a TRANSLATION instead of a transcript — "Вот вторая фраза" came back as "Here is the second phrase". Now, when it is not sure of the language and that language disagrees with the app's own, the phrase is asked again in yours. Confidently recognised foreign speech is left alone.
Live preview stopped costing more than the recognition itself.It re-read the text every half-second in the most expensive mode — twice the cost of the final pass over the same speech. Both now work the same way, leaving more of the graphics card free.
The app says plainly when it is running on the processor.With the Parakeet and GigaAM models this could happen: the graphics card failed to come up, recognition quietly moved to the processor and ran several times slower, while the screen still said "GPU". That fallback is now reported, along with what caused it.
You can see how long you actually waited.Settings → Performance now shows the time from releasing the key to the finished text, next to the engine's own time — the engine is only part of the wait.
The floating bar draws more cheaply.The level meter and the spinner no longer rebuild their gradients every frame, and a hidden bar no longer repaints on every window switch. It all looks the same.
0.14.0
A new business tier (Corp) — for teams.Server mode, where one powerful computer transcribes for everybody else, now belongs to it. An individual Pro licence still unlocks everything else: AI cleanup, translation, live preview, voice commands. A business key unlocks the Pro features too, on the computer and on the phone alike. Request it at sweetwhisper.app
0.13.2
Fewer stray repeats in recognition.When whisper doubled a short phrase (sometimes capitalising the second copy after a full stop — "…here. Here…"), the repeat now collapses; short human repeats are left alone.
Another doubling pattern caught— when whisper finished a phrase and re-emitted its BEGINNING at the end ("…with payment? Hello! Could we…"), that trailing echo of the opening is now trimmed.
The "Network sharing" tab is renamed to "Recognition server" and made clearer:the name now says what the mode does; a line up top states plainly that this computer is the server (node) and the consumer program is separate; the keys block notes that the key goes into the consumer settings and is shown only when issued.
Unicode text insertion is noticeably faster.Larger chunks and almost no delay between them (150 chars / 5 ms instead of 50 / 20) make long text insert several times faster. Both are now exposed under Settings → Text insertion; raise the delay if an app drops characters. The update applies automatically.
0.13.1
Short-clip recognition is noticeably faster.Re-enabled the optimization that processes a short recording in time proportional to its length — roughly twice as fast on short dictations, with the same text.
Removed the floating-bar white flashon the very first recognition after launch.
Fixed a rare doubled phrasein file and network recognition: the decoder sometimes repeated a whole sentence — the repeat now collapses.
Launching a game no longer throws a scary error.When the app unloaded the model for a game on the auto-unload list, a dictation attempt now shows a quiet "model paused" hint instead of a red "failed to load, check your GPU settings".
The Performance panel shows the last recognition time(previously blank); optionally, a per-stage timing breakdown in the log and the history.
The "Network acceptance" tab was reworked to be usable:an onboarding tour and a "how it works" card, clickable status rows (engine/device jump to Recognition), a Copy button on the address, and clear explanations of why key issuing is off on the trial and what "up to 3 keys" means.
0.13.0
Server mode: one strong machine transcribes for everyone (Pro).Flip the switch in settings — the app keeps working as usual and additionally starts accepting dictation from other devices on your network, transcribing it on its own GPU. A weak laptop no longer needs a model of its own.
Access by key only.Keys are issued in settings, shown once and revocable at any time; up to three at a time. The app listens on this computer alone — the mode is exposed through a tunnel, and no port is opened.
Your own dictation always comes first.Network requests wait in the queue; you do not. Start a game and the video memory goes to the game — the node answers "busy" rather than taking it back.
Somebody else's speech leaves nothing on your machine.Neither the text nor the audio reaches the history, the log or the disk.
A file the app cannot read no longer looks like a started job.It used to flash a progress tile before showing the error.
0.12.0
The voice-command journal is no longer on by default.It wrote what you said to disk, and it did so unasked for everyone who turned voice commands on. Off now, and turned on deliberately.
Dictation audio is no longer saved to disk by default.Recordings used to pile up in the data folder even when you had no use for them.
A custom model can no longer place files outside the models folder.The folder name was taken from the import as-is, and a crafted name wrote files where you never asked.
Progress spinners no longer freeze under "reduce motion".Every one of them: a motionless circle reads as "the app has hung", not as "work is happening".
The hallucination filter stopped cutting ordinary words."I have a bad cough" was inserted as "I have a bad" — the word vanished silently. The same happened to "laughter", "gunshots" and to the year "1988" inside a longer number. Such words are now removed only where the recogniser wrote them as a stage direction: in caps, in brackets, or as a line of their own.
"Stop" pressed right after recording starts is no longer lost.The bar already said "recording", the press disappeared without a trace, and you had to press again.
"Stop" on a file transcription no longer waits for the file to end.On a two-hour recording the button looked dead for the rest of the hour.
The splash screen stopped lying about progress.The bar crawled forward on its own, ran up to nearly half ahead of the truth, and the "still working" line appeared half a minute after everything had actually stopped.
A long recording through OpenAI no longer breaks off.The audio went out in one piece, past the service's 25 MB limit.
The app no longer declares itself English.A screen reader read the Russian interface in an English voice — sounds instead of words. The spell-checker takes its dictionary from the same place.
A language change reaches every window at once, not just the one you switched it in.
The offline app stopped calling home.Every window it opened dialled Google's font servers — for fonts that ship inside the program.
The "update channel is not answering" dot can now light up.The updater could stay silent for weeks and look perfectly healthy.
Microphone calibration can be reset.The numbers were shown with no way to revisit them.
The CapsLock preset stopped promising the impossible.It claimed the case would not toggle — and it did.
History tells "empty" apart from "could not be read".
The models folder is no longer swapped silently.If the folder you named is missing, the app says so instead of switching to its own and pretending all is well.
Colours and sizes were put right across all three themes.Eight passes over the screens: text legibility, touch-sized controls, keyboard operation, plain wording instead of technical.
The equaliser follows your accent colourinstead of staying purple forever.
Stray markup characters are gone from "What's new".
0.11.12
"Reset all settings" no longer took the history with it.The button promises to reset settings; the factory retention window came back along with them, and anyone who had chosen "keep forever" lost everything older than a month at the next sweep — silently, with no way back. The confirmation now names how many entries that costs, and says so plainly when the count could not be taken instead of staying quiet.
The cross beside an API-key field reported a removal that had not happened.When the credential store refused to delete the entry, the field on screen was cleared anyway while the key stayed put and kept going out with every request. The app now reads the store back and tells the truth.
Shortening the history window asks, from either screen.The same setting is reachable from two places and only one of them asked; the other deleted without a word.
A setting that failed to save no longer looks applied.The switch stayed in its new position, no error appeared, and the old value was back at the next open — which reads as the setting resetting itself. Fixed across every screen.
The context menu closes when you click away.It used to sit over the bar until you found something in it to press.
The interface scale applies to every window at once.It is changed in the settings window and took effect only there; the rest kept the old size until a restart.
The model benchmark stops inventing a reason."Accuracy is not shown because this language has no reference clip" also appeared when you used your own audio, and when every model failed.
Importing settings no longer reports losses that did not happen.Moving to a new machine told someone who had never entered an API key that twelve of them were skipped.
"Re-transcribe" in the history asks first.It rewrites the entry with no undo, while the delete button beside it confirmed twice.
Escape cancels a recording in every mode.
0.11.11
Carrying your data over from an older version can no longer be lost halfway.When the app failed to fetch your history and settings the first time — an unplugged drive, a file in use, a shutdown mid-way — the next start used to assume the move had happened and began from empty. It now remembers what it set out to do and finishes at the next opportunity. While a history is still on its way, no empty one is created in its place.
Settings are saved whole or not at all.A power cut or a failure during the write could previously leave a truncated settings file, and the app would start with half your preferences.
Running from a flash drive or a network share.Some filesystems cannot confirm a write the way an ordinary disk does. The app no longer treats that as a failure — while still refusing to pass over a write that genuinely was lost.
Crash reports arrive exactly once.A report can no longer vanish when delivery fails, nor be sent twice when it succeeds.
0.11.10
The portable version no longer loses its settings.Starting the app with a separate data folder could take the portable install's settings.json for itself and rename the original, leaving it to start with empty settings. Carrying old settings over now only copies them and never touches the source.
A missing settings file heals itself.If settings.json disappeared, the app used to keep running blind: nothing was ever saved and buttons quietly stopped responding. It is now recreated on the first attempt to read it, and the app stays usable.
The first-run wizard says so when it cannot save.The "Start using" button did nothing at all when the write failed. It now shows what went wrong and still lets you through.
0.11.9
The update check no longer waits for first-run setup to finish.The app used to start looking for updates only after you completed the wizard, so anyone who stopped partway would not hear about a new version even a week later. It now checks from the first launch, while the "update available" message still appears only after the wizard, so it cannot cover the wizard's own buttons.
0.11.8
The "Get Pro" button now opens the purchase page directly.It used to open the site's home page at the pricing block, so you had to scroll it, find another button there, and only then reach the form. It goes straight to the page that asks for your email now, and an English user gets the English shop rather than a translation of the Russian one.
The "your trial has ended" notice no longer vanishes after five seconds.It appears once in the app's entire life and is marked as shown immediately, so anyone who looked away at that moment simply found the AI features silently dead. It now stays until dismissed.
The gear can no longer stop opening Settings until you restart.If the settings window went away — its webview crashed, say — the app kept holding on to it and tripped over the same error on every click, which from the outside looked like nothing happening at all. It now notices the window is gone and builds a new one.
A rejected feedback submission no longer looks exactly like a delivered one: the form said "sent" whatever the server answered, and nothing was written to the log.
0.11.7
The About window no longer cuts off its own contents.Opened from the bar's menu it was noticeably smaller than when opened from Settings — two sizes because two parts of the app each built their own window. There is one window now, and it is tall enough to read the release notes without resizing.
The release notes in that window are shown in full again.Any bullet that ran over one line in the source text was truncated at the first — in every version at once, and the only clue was a sentence ending mid-thought.
0.11.6
The first dictation in the setup wizard no longer greets you with two errors.On the "try dictating" step the app was recording twice over: once for the wizard, once in the background. The second attempt died with a red "Engine error: Not recording", and the recognised text was also typed into whatever window had been open before the wizard — followed by a warning that it could not be inserted. The wizard now owns recording while it is on screen, and the text stays in its own field, as intended.
The live speech preview no longer pops up anchored to the hidden bar while the wizard is running.
0.11.5
A frozen floating bar can be revived. WebView2 sometimes stops delivering frames to the screen while everything inside keeps running: dictation works, but the bar shows a stale picture and looks permanently idle. The tray menu now offers "Restart the bar" and "Restart the app" — killing the process used to be the only way out.
Fixed the CUDA build (broken by a Rust update; it does not affect the shipped app, but the CUDA variant could not be built).
Website: dropped a VirusTotal verification claim that pointed at a different, older build; it now offers the release's SHA-256 to check against instead.
0.11.4
Short phrases in another language are no longer translated.Saying a couple of words in English used to come back as a Russian translation: “Yeah, that’s right” became “Да, это правильно”. The language is now detected from the speech itself, at any phrase length — not just long ones.
0.11.3
Updates are visible right inside the app.An available update now shows in Settings and the "About" window — not just on the tray icon. Every window shows it consistently, and you can install from any of them.
The "update available" notification no longer pops up over the loading screen at startup.
0.11.2
Clearer in-app updates.Installing an update now shows a prominent card with the download progress, status, and a notice that the app will restart. No more silent pause that looked like the app had frozen.
0.11.1
One-click update from inside the app.The "About" page now shows a prominent banner when a new version is available, with an "Update now" button — no more hunting for it in settings.
Onboarding: removed the stray lines around the theme swatches; the hotkey step no longer overflows off-screen when entering a custom combination.
0.11.0
A prettier, clearer first run.The setup wizard is redesigned: one segmented progress bar with the current step's name instead of competing dot rows, warm brand-style illustrations, and a compact color-swatch theme picker.
A clearer "Try it" step.The practice step now shows a live mini-panel — just like the real one — with a "Ready / Listening / Processing" status below it.
Clearer hotkey step.The key combination and the trigger mode are now split into separate, labeled sections, so the mode choice isn't missed.
Readability & accessibility.Text contrast meets WCAG AA on every theme (cream, dark, light), with balanced vertical rhythm on the short steps.
0.10.3
Friendlier installer on update.When SweetWhisper is already installed, the wizard now defaults to an in-place update ("Update (install over)") instead of uninstalling the old version first. Your settings, models, and history are kept.
0.10.2
Dictation can no longer wedge for good.If transcription gets stuck, the bar resets with a single button — no waiting, no restart.
Honest acceleration status.When the GPU is unavailable and transcription runs on the CPU, the app says so plainly ("CPU — GPU unavailable") instead of showing "Ready".
Text no longer lands in the wrong window.A cancelled or timed-out transcription won't paste its late result into whatever you switched to.
More reliable GPU startup.Third-party overlays (OBS, Wallpaper Engine, Steam/Epic overlays) no longer interfere with GPU initialization.
0.10.1
Routine security update. No functional changes.
0.10.0
Reworked first run.The wizard now recommends a model up front (GigaAM for Russian on a low-VRAM GPU) instead of making you choose blind, and the GigaAM catalog is complete. Irrelevant steps are skipped automatically.
No model is no longer a dead end.Dictating without a model opens the setup wizard; the same button appears in the empty model list.
First words are no longer clippedwhen you dictate right after launch. Releasing the hotkey before the engine is ready now cancels the take cleanly.
Less swallowed and duplicated speech:padding mode honours its own setting, and overlapping speech segments are no longer stitched in twice.
Short clips in "Auto" no longer hallucinate a foreign language.
The floating bar tells the truth:you can see when the model is unloaded, downloading, or failed; a hint when you speak too quietly; elapsed time in the tooltip. Errors go to an OS notification instead of breaking the pill.
Hotkeys:F5–F12 and the right Alt / Win keys are recognised; the trap where cancel was a subset of the record combo is rejected.
History:the player no longer offers audio that retention has already swept; search finds words with apostrophes and hyphens.
Settings are durable:every change is now flushed to disk, so a sudden power loss can no longer truncate or corrupt your settings file.
A transcription error stays until you dismiss it. The result window will not close while your cursor is on it.
Light and cream themes: warnings and errors are readable again (they used to wash out on the bright background). The Pro dialog's dimming scrim is back.
The 14-day Pro trial is announced up front instead of surfacing as a surprise.
Fixed broken GitHub links. "Open model folder" works on installed builds.
0.9.27
Model download source selector (Settings → Recognition): "Auto", "sweetwhisper.app mirror" (when HuggingFace is unreachable), or "HuggingFace only".
Changing the GPU device no longer reloads the model needlessly: when "Auto" and the selected device resolve to the same card, the reload is skipped; when the engine is pinned to another GPU until restart, an honest notice is shown instead of a long no-op reload.
Error reports are more useful: technical details (paths, model names) are no longer stripped from diagnostics, while private data is still scrubbed.
"Thank the developer" button in the About section.
Many small UI fixes: unified notifications, correct record-day date in stats, tidy wrapping of long model names, fixed invisible app chips, consistent terminology and copy.
0.9.26
Model download fallback: when HuggingFace is unreachable, models are fetched from the sweetwhisper.app mirror automatically.
"Save diagnostics file" button (Settings → Help): inspect the report contents before anything is sent.
Welcome screen after Pro activation: feature cards deep-link straight to their settings.
Polish: unified notification style in General settings, consistent "edition" terminology.
0.9.25
Reworked onboarding: per-step explanations and tooltips, plus a live press-to-test for your hotkey.
After first launch — a coach mark above the bar ("press your hotkey to dictate"); floating-bar buttons now have tooltips.
About section: "What's New" with version history.
0.9.24
Model library: "New" badges and clear speed/accuracy bars.
System theme (auto-follow Windows), clipboard-behaviour choice, history-retention presets.
Dedicated cancel-recording hotkey; active microphone name in the settings header.
Centered recording wave and an explicit "Stop" button in the floating bar.
0.9.23
Voice-commands overhaul: carrier verbs, friendlier input and key capture, export via dialog.
0.9.20
Heuristic recommendation that tunes the engine and model to your hardware (GPU + memory).
0.9.15
Cross-platform: Linux build and a unified cargo xtask workflow.
0.9.10
GigaAM Russian engine (Sber): automatic punctuation and capitalization.
Download the latest build on the home page. Already installed? The app will offer the update on its own.
