Changelog

VoiceWink has been in development since March 2026. These notes start at 1.23.280, released 30 May 2026. Releases before 1.87.382 were pre-launch builds.

1.87.382 2026-09-15

• First public release: the signed installer at voicewink.app, with the free 7-day trial on first run.

Pre-launch releases

1.86.381 2026-09-14

• About → Support: a "Suggest an improvement" button opens a pre-addressed email.
• Report a problem: the email ends with the VoiceWink version.
• About: a "Third-party licenses" link opens the bundled licence notices.

1.85.380 2026-09-13

• A model name typed into the AI Enhancement model box no longer appears in VoiceWink's own log lines unless it looks like a model id; provider error responses and the opt-in full prompt log are unchanged.
• Word Replacements: a rule switched off dims immediately instead of after leaving the page.
• Problem reports that include logs now carry a redacted prompt summary (text replaced by its length); the full prompt log stays opt-in.
• The setup wizard's "Help improve VoiceWink" choice now also switches the full prompt log on or off; image prompts stay a separate opt-in in Settings → Diagnostics.
• Word Replacements: the edit row keeps its buttons on the card and the list shows the whole original before trimming.
• Word replacements can be edited in place: the pencil on a row loads it into the boxes above, then Save or Cancel.
• App Modes keep the enhancement you chose for them; it no longer resets to the default when VoiceWink starts.
• The Dictionary page's two descriptions are shorter.
• The Cerebras model list no longer offers gemma-4-31b, which Cerebras stopped serving.
• Filler-word removal no longer deletes real words from Dutch, German and French dictation ("er", "eh", "ah"), and is skipped entirely when a non-English language is selected.
• History is no longer deleted automatically after 7 days by default; the auto-delete option is still available in Settings.
• "Reset all settings" now says what it keeps, and restores Right Ctrl (not Right Alt) as the recording hotkey on keyboards that use AltGr.
• The setup wizard's summary now lists minimize-to-tray, crash reports and the update choices, and says when AI enhancement is on without an API key or model.
• The setup wizard's hotkey step explains Instant recording, and "Send Enter after paste" now says that it sends the message in chat apps.
• Rerunning the setup wizard now updates the AI Enhancement switch on its page.
• "Verbose debug logging" in the Log Viewer now writes debug-level detail to the log, and covers startup too.
• The log opens each session with the app version, Windows version, architecture and the main settings, and problem reports include the same summary.
• Deleting all history no longer deletes the log files.
• The crash-report choice takes effect immediately, in the setup wizard and in Settings.

1.84.379 2026-09-13

• The Dictionary starts with VoiceWink's own names — the product, the providers, the local models — and three spelling fixes ("voice wing" → VoiceWink, "DeepGram" → Deepgram, "parkit" → Parakeet); remove any you don't need and they stay removed.
• The AI-enhancement model list no longer offers OpenAI's gpt-live-1, a voice-session model that cannot enhance text.
• OpenAI's gpt-image-2.5 Flare and Sunburst now offer the 1K/2K/4K sizes and the 2:1, 1:2 and 9:21 ratios, like gpt-image-2.
• Image options now use each provider's own names: quality reads Low / Medium / High / XHigh / Max (XHigh and Max are new, gpt-image-2.5 only; the old Maximum tier is now High and still sends the same value) and the smallest size reads 512 instead of 0.5K.

1.84.378 2026-09-13

• Public releases now have a page on GitHub with these notes and the signed installer to download.
• Home shows how long your free trial has left; the last three days turn amber.
• After the free trial ends, the first blocked recording opens the License page for you, and clicking the recorder pill opens it.
• AI enhancement asks reasoning models to answer directly instead of thinking at length, which ends the occasional "No answer: token limit" on Cerebras.
• The re-enhance dialog suggests Qwen 3.8 for Cerebras; Gemma 4 no longer works there, so pick another model if you had it selected.

1.83.377 2026-09-11

• Send Enter after paste is on by default; switch it off in Settings if you want the caret to stay on the line.
• The License page now names this device, so support can free the right seat when a device dies.

1.82.375 2026-09-11

• Your license key is re-checked online once after this update, before recording continues.
• Send Enter after paste is now off by default; switch it on in Settings if you use it for chat apps.

1.80.373 2026-09-09

• Clearer message when an AI model stops before answering.
• License page: the key box shows your stored key (masked) instead of a sentence above it.

1.79.372 2026-09-09

• Light theme: amber warning and advisory text across the app is now readable (it was too faint).
• License page: when a license has been disabled, the page shows which (masked) key it was.

1.78.371 2026-09-08

• Privacy Policy: the third-parties section no longer reads as if license checks only happen with cloud features (you will be asked to accept it once).
• Restoring the window no longer adds a second taskbar icon beside the pinned one (if you still see two, unpin and pin it again).
• License page: the check button is a refresh icon beside the status, and the result shows on that line without moving the card.
• A disabled license's status chip reads DISABLED.
• Setup wizard: shorter free-trial text; License page: Start free trial sits between Activate and Buy, and the trial length is stated once.

1.78.370 2026-09-07

• The license is also re-checked with the server about once a day while VoiceWink stays open.
• The free trial is 7 days, and deleting the app's data no longer restarts it.
• Activating a key does not end the free trial; deactivating returns to the days that were left.
• After a check, the License page says what the license is now.
• "Check now" is the first button on every License panel, and button rows wrap instead of clipping in a narrow window.
• Starting the free trial in the setup wizard takes one click after a failed key entry.
• A recording blocked after the trial says the trial has ended.

1.77.369 2026-09-06

• The license is re-checked with the server on every start after setup is complete.
• The free trial no longer needs a key or an email.

1.76.368 2026-09-05

• A license key disabled in the Lemon Squeezy dashboard is now recognised on the next check instead of staying active.
• The License page no longer carries a refused activation's warning onto the tryout card, and a failed check now reads in red.
• The License page's key form is one row of buttons, with the free-trial-key hint always shown.
• License warnings and the status chip are readable in the light theme.

1.75.367 2026-09-05

• On a PC with two graphics cards, local models now use the dedicated one instead of whichever Windows lists first.
• The GPU acceleration row names the graphics card the selected model is running on.

1.74.366 2026-09-04

• "Check now" on the License page now tells you what happened, including when it couldn't reach the license server.
• The License page keeps itself up to date instead of showing a state the app has already moved on from.
• Trying VoiceWink without a key is no longer offered on a machine that is already doing so, and that screen now has a way back.
• An expired license key says when it expired, and the status badge agrees with it.
• Several License page messages no longer give a reason the app cannot actually know — being offline, needing to reconnect, or why a key was refused.
• Smaller License page fixes: the disabled-key screen can look up a lost key, the device count no longer reads zero on the machine you just activated, and every Buy button shows where it goes.

1.73.365 2026-09-04

• Local models switch to the processor when the GPU check finds your graphics card too slow.
• The GPU acceleration row confirms when the selected model is running on your graphics card.
• Whisper no longer transcribes on a graphics card that failed its check; the recording is kept so you can retry on another model.
• The GPU acceleration warning now shows only when it affects the model you have selected.
• The GPU acceleration row says when your graphics card is being checked, and clears its warning the moment you switch it off.

1.72.364 2026-09-04

• The GPU acceleration switch moved from Settings to the Models page, above the local models it speeds up.
• Setup wizard: the license step keeps its layout when you go back, leads with trying VoiceWink without a key, and shows a Buy button; an expired trial key now leads with Buy on the License page.
• Flipping GPU acceleration now offers to restart VoiceWink so the change applies right away; a restart is refused while a recording or transcription is running.
• If you have crash reporting switched on, a failed graphics-card check now reports your graphics adapters and their driver versions, so a faulty driver can be identified and worked around.

1.71.363 2026-09-03

• GPU acceleration is checked with a short test clip at start-up; a graphics driver that produces no text is bypassed and local models run on the processor instead.
• A local model that returns no text on the graphics card is retried on the processor, which VoiceWink then keeps using until GPU acceleration is switched off and on again.
• A short recording that captured only silence is treated as no speech instead of restarting the microphone stream; only repeated silent recordings still trigger the restart.
• After a graphics driver update, GPU acceleration is tested again at the next start on its own, instead of waiting for the next VoiceWink update.

1.70.362 2026-09-03

• Report a problem: if your email app rejects the report, you can now try again before attaching the file yourself
• The log now records which GPU each local speech model is running on
• The GPU acceleration setting shows off and greyed out, with the reason, on a PC that has no GPU driver VoiceWink can use
• Settings now says when local Whisper models are running on the CPU because GPU support could not start on this PC

1.69.361 2026-09-02

• Models page: local models now show GPU speed stars when GPU acceleration is active, CPU speed stars otherwise
• Setup wizard: the license step now separates the free 7-day trial key from trying without a key, and no longer mentions signing up

1.68.360 2026-09-02

• Local Whisper models now use your graphics card (GPU) when available — much faster on capable machines, automatic fallback to CPU otherwise
• Parakeet transcription now also uses your graphics card when available — several times faster decoding, automatic fallback to CPU otherwise
• The first dictation after installing no longer pays a one-time GPU preparation delay — it now runs in the background at startup
• New setting: GPU acceleration (on by default, under Settings → General; takes effect at next start)

1.67.359 2026-09-01

• Fixed Parakeet sometimes returning nothing for a very short dictation
• Parakeet transcribes a bit faster on PCs with more than four processor threads
• Whisper Large V3 Turbo now shows the same accuracy rating as Medium
• The larger local Whisper models transcribe noticeably faster
• Fixed local Whisper occasionally repeating a sentence
• Parakeet now shows a five-star speed rating

1.66.358 2026-08-31

• On keyboards where Right Alt types characters (AltGr), the setup wizard now suggests Right Ctrl as the recording hotkey
• Hotkey menus and warnings now use friendly key names like "Right Alt" and "Ctrl + Space", with Space, letters and digits in their own sections
• Local Whisper models now show a more realistic speed rating on the Models page
• Accuracy ratings on the Models page now come from real measurements - two models changed rating
• Speed ratings for cloud models now come from real measurements - they all rate equally fast
• Steadier recording waveform: the bars no longer go flat in the pauses between words.
• Hotkeys can now be two-key combinations such as Ctrl+Space, so AltGr stays free for typing € and @.

1.65.357 2026-08-29

• Fixed: a Bluetooth headset could stay muted for calls after dictating

1.64.355 2026-08-29

• Speech-model downloads now come from VoiceWink’s own server; the privacy policy was updated to match, so you’ll be asked to accept it once
• Fixed: very quiet recordings could come back empty from a cloud transcription provider
• Speaking faintly against loud background noise is less likely to be discarded as “No speech detected”

1.63.354 2026-08-26

• Quiet dictation is no longer thrown away — speaking softly, or with your microphone turned down, used to come back as “No speech detected”
• The recording waveform now follows your voice at any microphone level: it rests when you are not speaking, moves while you are, and stays small when the level is too low to transcribe
• Fixed: a microphone that stopped delivering sound was reported as “No speech detected”; VoiceWink now says so plainly and recovers on its own
• Fixed: the “selected mic unavailable” warning was hard to read on top of the live waveform

1.62.353 2026-08-25

• The old Parakeet engine's leftover files are now removed automatically once the new engine has delivered a transcription
• Fixed: sending a problem report could close VoiceWink instead of opening your email
• Fixed: the faster Parakeet engine never started on installed builds — dictation quietly used the older, slower one, and a brand-new install could not transcribe locally at all

1.61.352 2026-08-24

• Parakeet transcription is sharply more accurate — far fewer dropped words and sentences; the model updates itself in the background (a one-time 940 MB download), and the Models page then offers to remove the old files
• Updated privacy policy wording about local speech-model downloads; you'll be asked to accept it once

1.60.351 2026-08-23

• The recorder pill now clears after 10 seconds instead of 15; errors still stay until you dismiss them
• Fixed the taskbar icon turning into a blank Windows icon after reopening the window from the system tray
• Corrected the Lemon Squeezy name in the app and legal documents; the updated legal text asks for one re-acceptance
• Fixed dictated text being pasted twice into slow-responding apps

1.59.350 2026-08-22

• Updated license agreement and privacy policy (now lawyer-reviewed); you'll be asked to accept them once
• Starting an image generation now shows its progress right away instead of waiting behind a retry message
• History keeps your place when a new transcription arrives instead of jumping back to the top
• Report a problem now explains it when your email app can't take the attachment automatically

1.58.349 2026-08-17

• The prompt trace now shows which language was used and why

1.55.345 2026-08-17

• Redo now pastes into the text box you dictated into, instead of falling back to the clipboard
• Redo no longer presses Enter, so a corrected message is never sent without you reading it
• The microphone picker in Settings no longer squeezes its description into a narrow column
• Dictation started while a model or retry dialog is open is copied to the clipboard instead of being typed into VoiceWink

1.54.344 2026-08-15

• Parakeet no longer drops the final word of some dictations

1.53.342 2026-08-13

• The mini recorder now says "Enhancing…" for the whole enhancement, instead of swapping in a stop-button hint partway through.

1.53.341 2026-08-10

• The mini recorder now says "Done" when your text or image is pasted.

1.53.339 2026-08-09

• VoiceWink can now install updates on its own — on by default, and never while you are recording, transcribing, or generating an image.
• Onboarding now asks whether VoiceWink should check for and install updates automatically.

1.52.337 2026-08-08

• Quiet audio files now get the same volume boost as quiet recordings before transcription.
• Download and Audio Transcribe buttons now turn into Cancel while working, instead of sitting greyed out beside a separate cancel button.
• Onboarding no longer shows the Terms and Privacy step once you have already accepted them.
• Onboarding now offers local transcription first and selects it by default.
• Onboarding's startup step now also offers "Minimize to tray".
• Transcription model names appear in plain language; an unrecognised selection now says so instead of showing a file name.
• Onboarding's download step now explains its recommendation for your chosen language.
• The unsupported-language warning no longer suggests auto-detect will cope, and during onboarding it names the model to download.
• Fixed: starting a recording twice in quick succession could silently record nothing.
• The recorder now warns within seconds if the microphone stops mid-recording.
• Fixed: a recording started right after a slow one could be cut short.

1.51.336 2026-08-06

• Fixed: long recordings could stop partway through when transcribing with a local Parakeet model.
• Fixed: quiet recordings could transcribe as empty or incomplete with the Parakeet model.
• Fixed the app taking minutes to become responsive when launched right after a Windows restart.

1.50.335 2026-08-05

• Recording now starts instantly when you press the hotkey. Settings → Audio → "Instant recording" (on by default) keeps the microphone ready; turn it off to release it between recordings.

1.43.334 2026-08-05

• Local models now say which engine they run — "Parakeet" and "Whisper Tiny/Base/Small/Medium/Large V3 Turbo" — with Parakeet first and the Whisper models largest first.
• Recordings are filtered before being converted, so high-frequency sound no longer folds down into the speech range.
• Settings is reorganised: a General section at the top, Data Management split into History cleanup, Your data and Settings backup, and the prompt, recording and crash-report options moved to a new Diagnostics section — so the Log Viewer shows the log again.
• The image model list opens instantly, plus smaller fixes and polish.

1.42.333 2026-08-04

• Parakeet (Fast Multilingual) — a new local model, roughly 15× faster than the previous default on CPU. New installs now start on it.
• The English-only local models are gone; if you used one you now use its multilingual version. Local model files VoiceWink no longer supports are deleted, freeing the space they used.
• The Models page now tells you when the selected model cannot transcribe your chosen language, instead of silently detecting it.
• AVIF reference images, a fixed crash in the image options dialog, correct clipboard copying for non-PNG images, and smaller display fixes.

1.41.332 2026-08-03

• OpenAI GPT Transcribe — more accurate than GPT-4o Transcribe and cheaper per minute. Your dictionary and trigger words can be sent to it for better recognition.
• Compact versions of the Small, Medium, Large V3 and Large V3 Turbo local models — same models, roughly a third of the download.
• Choose which microphone VoiceWink records from (Settings → Audio); if it is unavailable, the system default is used for that recording.
• Very soft and whispered recordings are boosted automatically so speech recognition can hear them.
• Cleanup prompts are much better: misheard words are corrected, spoken punctuation becomes real marks, numbers and dates are written normally, long dictation comes back in paragraphs, and nothing is filled in with placeholders like "[Name]". The E-mail prompt adds a greeting and closing only when you actually wanted one.
• A dictation that fails while connecting on a weak network now retries once automatically instead of asking you to tap Retry.
• The recorder shows "Starting…" while the microphone opens, and recording begins a little sooner.
• Every cloud provider now has a "Get API key" link, model lists carry short guidance, and the mini recorder's colours match its state — plus smaller fixes and polish.

1.34.331 2026-07-31

• The recorder now tells you when enhancement could not run because no model or an unsupported model was selected, instead of quietly pasting the raw transcription.
• Fixed a crash after saving an API key or switching AI providers quickly.
• Tidier Gemini, OpenRouter and Cerebras model lists, with Gemma models now offered.

1.34.330 2026-07-30

• Paste reliability: dictated text now lands in Claude and WhatsApp instead of silently going nowhere, VoiceWink checks that the text actually arrived and retries when it did not, and a paste that fails shows a warning instead of reporting success.
• Switching to another text field at the last moment no longer pastes into the wrong one, and an unresponsive app can no longer grab focus seconds after a paste.
• Right-click the mini recorder to hide it while a long job finishes; restore it from the tray.
• The mini recorder is kept out of screenshots and screen recordings by default.
• A rejected license check now stops recording instead of continuing on the last successful check.
• Clearer license wording, refreshed default models, and smaller fixes.

1.34.329 2026-07-27

• Prompt and output logging is now available in every build (off by default), and problem reports can include prompt logs and kept audio recordings.
• New option to keep audio recordings for 7 days for debugging.
• Clearer provider error messages in the mini recorder.

1.33.328 2026-07-26

• Set the speech recognition language per AI enhancement — a translation prompt can carry its own source language.
• Mini-recorder messages now say what happened to your text ("Text pasted", "Enhancement timed out — pasted transcription").
• The Models page notices Whisper files deleted from disk instead of pretending they are still installed.
• The delete-history dialog explains exactly what is removed and what is kept.

1.33.327 2026-07-25

• Transcription errors now name the real cause instead of blaming your connection or your API key.
• Shorter, plainer status messages in the mini recorder.

1.33.326 2026-07-24

• Smaller fixes: pages no longer jump back to the top when a button refreshes them, the amber Retry pill no longer reappears endlessly, and no-speech detection is hardened against a rare crash.

1.33.325 2026-07-24

• Near-silent recordings no longer paste hallucinated text ("Thank you.", etc.) — real voice-activity detection screens recordings, and a blocked recording offers Retry instead of being discarded.
• Export and import your settings to a file (Settings → Settings backup), excluding API keys, license and device-specific state.
• The default local transcription model is now Small (was Base) for better accuracy.
• A new Enable Dictionary switch, a choice of whether minimizing hides to the tray, and smaller fixes.
• Updated the bundled SQLite database engine to fix a security vulnerability (CVE-2025-6965).

1.33.324 2026-07-19

• Deleting an enhancement that an App Mode uses now warns you first and names the affected App Modes.
• Adding a prompt or an App Mode opens its editor immediately instead of appending a blank one to the list.
• Updated builds refresh the built-in prompts and App-Mode presets when they have been improved; your own custom prompts are kept.
• Smaller fixes and polish.

1.33.323 2026-07-18

• Prompt trigger words are recognised far more reliably, including when the transcription adds punctuation ("Hey, assistant.").
• Many more Dictionary words now reach the recognizer in each dictation.
• New application icon, plus smaller fixes.

1.32.322 2026-07-18

• Dictionary words now improve Deepgram transcription automatically, and ElevenLabs transcription via an opt-in toggle (ElevenLabs bills keyterms at +20%).
• "No speech detected" now keeps the recording and offers Retry — switch the transcription model and try again instead of losing the dictation.
• Dictations keep their language through enhancement unless the prompt itself translates, and cleanup prompts no longer answer questions you dictate.
• Improved built-in prompt templates, and ElevenLabs Scribe v1 selections now use Scribe v2.

1.31.321 2026-07-17

• AI enhancement no longer pastes a reasoning model's internal thought process.
• Push-to-talk no longer stops working after a recording is stopped from the pill, tray, or main window.

1.31.320 2026-07-17

• Various improvements and bug fixes.

1.23.284 2026-06-13

• Export my data, Delete transcription history, and Delete all my data (Settings → Data Management).
• In-app About card and a "What's new" dialog after updates.

1.23.283 2026-06-12

• Starting a free trial no longer leaves the License page stuck on the key-entry form.

1.23.281 2026-06-03

• VoiceWink checks for updates after launch and daily, with a dot on the Updates sidebar entry.

1.23.280 2026-05-30

• Internal release-pipeline, installer, and packaging improvements.

Development started 2026-03-03

• First commit in the VoiceWink repository.
• Release notes were not kept before 1.23.280.