PrivateTranscribe
PrivateTranscribeChangelog

Changelog

New features, improvements, and fixes for PrivateTranscribe.

v0.24.2October 24, 2026

Saved Keys and History Kept Safe, and Steadier Long Dictations

  • FixedSaved API keys are no longer lost when Windows cannot read the key file at start. Before, the app wrote an empty key file over it a few seconds later. Now the file stays until you save a key, and the next start reads it again. If you save a key first, the old file is kept beside the new one.
  • FixedHistory is no longer cut to 50 dictations when the app's window settings are lost, and the dictionary is no longer replaced by the default list.
  • FixedImporting a settings file never deletes history. A file with a lower history limit than this PC's is skipped, and the import summary says why. The import now also takes a history limit above 10,000, and an idle shutdown of 0, which keeps the local speech engine running.
  • FixedAn imported dictation hotkey works at once. Before, the old hotkey kept starting dictation until a restart.
  • FixedA dictation hotkey that Windows refuses is not kept, and the old one keeps working. Before, a refused key such as "-" could take over that key in every app while the old hotkey stopped working.
  • FixedA microphone that never opens shows "Microphone Not Responding" after 15 seconds. Before, the hotkey did nothing until the app was restarted.
  • FixedA long dictation no longer stays on processing forever when its recorder stops answering, which can happen when the microphone is unplugged.
  • FixedA silent minute in a long dictation adds nothing to the text. Before, Whisper could put an invented "Thank you." or "[Music]" in the middle of the text, the app showed "No Audio Detected", and with cloud transcription the minute was sent four times and marked as missing.
  • FixedA minute of a long dictation that fails to transcribe is retried, and marked once with "[... missing section ...]" if it still fails. Before, a minute local Whisper could not transcribe could vanish with no mark, and the mark could come out broken or twice.
  • FixedThe last retry of a long dictation keeps the language the dictation started in.
  • FixedA dictation you cancel with Escape is never sent to OpenAI. Before, with the OpenAI fallback turned on, a canceled dictation whose local transcription then failed was still uploaded.
  • FixedWhen the paste helper hangs or crashes after it has pasted, the app no longer asks you to press Ctrl+V, which pasted the text twice.
  • FixedLocal Whisper is no longer marked as stopped after one slow health check. Before, the next dictation restarted it and waited for the model to load again.
  • FixedA recording the app cannot convert shows "Audio conversion failed" instead of FFmpeg's raw output.
  • FixedTemporary audio left behind by a crash is removed the next time the app starts. Only files the app made itself and that are more than a day old are removed.
  • FixedThe local AI model no longer loads at every login while AI enhancement is off.
  • FixedInstalling an update installs the version the app says is ready. Before, an older downloaded update could be installed after a newer one was found, and a pulled update could still install itself when you quit.
  • FixedCorrection Memory again offers to learn from a correction you copy after a dictation. Before, the offer never appeared.
  • Fixed".gitignore", ".NET" and other words that start with a dot keep the space in front of them.
  • FixedRepeated words are collapsed only as whole words. Before, "the the theory" could come out as "theory".
  • FixedDictionary words that start or end with a letter such as ø, æ or å, like "Østerbro", now get the dictionary's capitals, and "tuborgøl" no longer turns into "tuborgØl". Words such as ".NET" and "C++" are matched too.
  • FixedParakeet leaves no stray "?" or "!" where it removed a filler word. Before, "Hmm? I don't know." pasted as "? I don't know."
  • FixedChinese, Japanese and Thai dictations of 20 seconds or more no longer get a false "may be incomplete" warning.
v0.24.1October 20, 2026

AI Cleanup That Never Holds Up a Dictation

  • FixedAI cleanup can no longer hold up a dictation. If the AI provider does not answer in time, your words are pasted as you said them. Before, a provider that never answered kept the dictation waiting, and every later cleanup that session was skipped.
  • FixedA wrong or expired API key fails at once. Before, it was tried three more times, which added about 7 seconds to every dictation.
  • FixedA long dictation is no longer cut short by AI cleanup. If the AI's reply stops at its length limit, your dictation is pasted without cleanup instead of the shortened text.
  • FixedAI cleanup works with Claude Sonnet 5 and Claude Opus 5.
  • FixedA custom AI endpoint whose model has the same name as another provider's model no longer skips cleanup.
  • FixedA Windows volume of 0% stays at 0% after a dictation. Before, it came back at 100%.
  • FixedPressing Escape during a recording puts back the volume, paused media and your voice call's mute key.
  • FixedA quick tap of the dictation hotkey no longer keeps your voice call muted for up to 10 minutes.
  • FixedIf the app closes or crashes while it holds your call's mute key, it lets go of the key the next time it starts.
  • FixedSmart Context works with Parakeet and cloud transcription. It reads the window when the dictation starts, so the text comes from the window you dictated into. Before, only local Whisper got the window text in time.
  • FixedSmart Context names the real app in front, so editors are recognized for active file context and password managers are recognized by their program name.
  • FixedWindow titles with ø, æ and å reach AI cleanup as written.
  • ImprovedSmart Context removes more secrets before any text goes to AI cleanup, including whole provider API keys, private keys, passwords in config files, XML and command lines, and password labels in Danish, Norwegian, Swedish, German and French. If removing them fails, none of that text is sent.
  • ImprovedSmart Context reads only the file your editor's title names, and never a file whose name says it holds secrets or credentials. The part of the file it reads goes through the same redaction as the window text.
  • ImprovedSmart Context reads nothing from more password managers, such as Keeper and a browser's saved passwords page.
  • ImprovedThe local Whisper log no longer records the words from your dictionary or the file names Smart Context found, only how long that text is.
v0.24.0October 20, 2026

Speaker Labels Fixed and Read Aloud on the Overlay Button

  • NewWhile Read Aloud speaks, the round overlay button shows bars that move with the voice, the same way it shows your voice while you dictate. The bars rest while a read is paused, and the button goes back to the logo when the read stops.
  • ImprovedOn PCs with 3 or 4 cores and no hyper-threading, Parakeet leaves one core free, so one busy background program no longer slows a dictation down as much. In our tests it made Parakeet 1.6 to 1.9 times slower before, and now at most about a third slower.
  • FixedSpeaker labels work in the installed app. In 0.23.0 they failed every time.
  • FixedIf speaker labels fail, the Transcribe page keeps the transcript and shows it without speaker names, with a note that says why. Before, the whole transcript was lost.
  • FixedLocal Whisper works when your Windows user folder has letters such as ø or æ in its name, or letters from another alphabet such as Greek or Chinese.
  • FixedParakeet and speaker label models download and install when your user folder name uses letters from another alphabet, such as Greek or Chinese.
  • FixedAuto-start can be turned off when your user folder has letters such as ø in its name. The switch shows its real state, and changing it no longer reports a false security software error.
  • FixedDictation records from the Windows default microphone while the microphone you picked is unplugged, as the microphone picker says, instead of stopping with an error. With no microphone at all, the error says no microphone was found.
  • FixedEvery key the hotkey picker offers works, including the arrow keys, Enter, Print Screen, the number pad and media keys, alone or with modifiers. Before, such a hotkey could be saved but never started dictation.
  • FixedHotkeys made only of modifiers, such as Ctrl+Shift, start dictation with the left or right keys, pressed in any order. They start once the keys are held alone for a moment, so another shortcut such as Ctrl+Shift+T, a mouse click or drag with the keys held, or typing with AltGr does not trigger them.
  • FixedAfter Ctrl+Alt+Del, a Windows permission prompt or locking the PC, a key you let go of there no longer counts as held, so a hotkey no longer fires on the wrong keys.
  • FixedChanging the dictation language, Translate to English or the custom transcription address applies to your next dictation, without a restart.
  • FixedReset app data deletes your saved API keys, history, settings and logs. It restarts the app to finish the job and lists anything it could not delete. It removes only our own folders, never a linked folder or the Whisper models other programs keep. Before, it reported success while some of your data stayed on disk.
  • FixedConverse no longer freezes the whole app when an answer contains a numbered or lettered list.
  • FixedConverse no longer repeats a sentence after a pause, such as while the agent runs a tool, and no longer stays on speaking after the last sentence.
  • ImprovedConverse and Agent mode use your own Claude Code sign-in and setup, such as Bedrock or a gateway, the way Claude Code does in a terminal. The Anthropic API key saved in the app is no longer used for them.
  • FixedThe suggested local Ollama address can be saved as a custom AI endpoint, and an endpoint that needs no API key works without one. Custom addresses accept plain http on this PC and https for any host, and a refused address shows why under the field.
  • FixedYou can type an address into the custom transcription Endpoint URL field. Before, only a pasted address was kept.
  • FixedLocal AI cleanup loads its model once with room for any dictation. A long dictation no longer makes the model load again, and the memory it reserves stays small.
  • FixedThe Read Aloud word highlight hides when another window covers the app being read, and comes back with that app.
  • ImprovedAPI keys saved in the app stay inside the app. Programs it starts, such as Claude Code in Converse and the commands it runs, do not receive them. A key you set in Windows yourself is no longer replaced by the saved one.
  • ImprovedWith the custom AI endpoint selected, dictation text goes only to that endpoint, whatever the model is called. With custom transcription selected, audio goes only to the address you entered. If the address cannot be used, the app stops with the reason and sends nothing to another provider.
  • ImprovedTemporary dictation audio always stays in your own Windows temp folder, also when its path has a space or letters such as ø. At startup the app deletes audio that older versions left in a shared folder.
  • ImprovedThe whisper.cpp and llama.cpp engines in the installer come from fixed versions that are checked against a SHA-256 fingerprint when the app is built. The push-to-talk key listener is built only from the app's own source code.
v0.23.0October 17, 2026

Cleaner Long Files and Faster Long Dictations

  • ImprovedLocal Whisper now cuts a file at pauses in speech into pieces of up to 30 seconds and sends only the speech. Files no longer pick up invented sentences in quiet stretches or drop spoken ones, and timestamps and speaker labels keep their place. In our tests on 10 minutes of Danish speech, wrong words fell from about 40% to about 15%.
  • ImprovedThe Transcribe page shows a progress bar for most files, because each finished piece is a real step.
  • ImprovedDictations longer than a minute are transcribed while you talk instead of after five minutes, so stopping leaves only the last part to wait for. With Whisper Turbo on a PC without a graphics card, the wait after a two-minute dictation fell from 17-36 seconds to 5-9 in our tests.
  • FixedA dictation is only split where you pause for at least 0.6 seconds, so a cut no longer lands on a breath in the middle of a sentence.
v0.22.0October 11, 2026

Parakeet for PCs Without an NVIDIA Card

  • NewParakeet is back as a local dictation engine. It now runs inside the app and opens no network port.
  • NewSetup picks Parakeet on a PC without an NVIDIA card, or with one under 6 GB of video memory, when the PC has at least 4 cores, AVX2 and 8 GB of memory and every language you speak is one Parakeet supports. A 10-second speed test runs after the download, and if Parakeet is too slow, setup deletes it and goes back to Whisper.
  • NewThe Dictation page lets you choose Parakeet or Whisper, and says why when your PC cannot run Parakeet. If you use Whisper on a PC that suits Parakeet, a one-time card offers the switch.
  • NewWhen a Parakeet dictation fails, the same recording is transcribed with an installed Whisper model instead. Cloud transcription is tried only if you allowed it and no local Whisper model could run.
  • ImprovedParakeet dictation leaves out hesitation sounds such as "uh". Sounds that are real words in a language you speak, such as Danish "er", stay.
  • ImprovedParakeet's language list names the 25 languages the model supports, Danish included.
  • ImprovedIf you used Parakeet before 0.21.0, the app moves you to Whisper once. You can pick Parakeet again on the Dictation page.
  • ImprovedWhen the app cannot read your hardware, setup picks Whisper Base instead of Turbo, which is slow without a graphics card.
  • FixedParakeet and speaker detection models install correctly on PCs with Git for Windows on the PATH.
  • FixedA model download that gets a 404 no longer leaves an empty temporary file behind.
  • ImprovedRead Aloud ships with sharp 0.35.5, which fixes the libvips and libheif flaws in the version it used before.
v0.21.1October 8, 2026

Read Aloud Clipboard Fix and Security Updates

  • FixedRead Aloud puts back everything you had copied, including images and formatted text. It used to keep only the plain text.
  • FixedSettings backups include your Read Aloud voice and reading speed and the call-mute shortcut and switch.
  • FixedImporting settings no longer shows Import Failed after the import worked.
  • FixedScreen readers announce the name of every settings switch and every dictionary remove button.
  • ImprovedThe app runs on Electron 41.10.6, up from 41.5.0, which fixes several Electron security issues.
  • ImprovedThe tar, js-yaml, ws, form-data and electron-updater libraries the app ships with are updated to versions that fix published security advisories.
v0.21.0October 6, 2026

Open Source on GitHub

  • LaunchPrivateTranscribe is open source under GPL-3.0-or-later, and the source code is on GitHub.
  • ImprovedConverse, Agent mode and the Action Engine are behind a new Experimental features switch in Settings, off by default. If you use them, turn the switch on and your saved settings for them come back.
  • ImprovedEvery Whisper, Parakeet, Read Aloud, speaker detection and local AI model download is checked against a fixed SHA-256 fingerprint before it is used. A file that does not match is deleted.
  • ImprovedCopies built from the source code never check for updates and send no usage analytics. The official installer is unchanged.
  • ImprovedThe app uses the Manrope font.
  • NewCoding prompt instructions you can edit, with save and reset, kept separate from the writing style.
  • FixedLocal Whisper dictation no longer splits words with stray spaces.
  • FixedCoding enhancement with a local model keeps to the words you said instead of writing code of its own.
  • FixedThe microphone meter recovers when it freezes during startup.
  • FixedSaving settings keeps lines you added to the .env file yourself, such as the debug log level.
  • FixedLowering the music while you record works when your Windows profile path contains a curly quote.
  • FixedGPT-OSS 20B downloads again.
  • FixedThe local AI server and the local Whisper server answer only the app. Each start makes a new random key or path, so a web page can no longer reach them.
  • FixedThe retired Parakeet server can no longer start. It listened on every network interface without a password.
  • FixedSmart Context takes a file name only from a known code editor's window, so a web page title cannot choose which file is sent to a cloud model.
  • FixedThe app opens only its own page as a local file, and a file dropped outside the Transcribe drop zone is refused instead of opened.
  • FixedConverse starts Claude Code with only your own Claude Code settings. A project's hooks, MCP servers, skills and instruction files load only after you choose to use this folder's setup, and only while those files match what you were shown.
v0.20.2September 29, 2026

GPU Transcription Without an Extra Install

  • FixedGPU transcription starts on PCs without the Microsoft Visual C++ Redistributable, instead of falling back to the CPU. The app now places the Microsoft runtime files the GPU engine needs beside it, so an engine you already downloaded works without downloading it again.
v0.20.1September 29, 2026

No Extra Install for Parakeet and Read Aloud

  • FixedParakeet transcription and Read Aloud start on PCs without the Microsoft Visual C++ Redistributable. The app now ships the Microsoft runtime files both engines need, instead of relying on a copy another program installed.
v0.20.0September 29, 2026

Free for Everyone, No Word Limit

  • LaunchPrivateTranscribe is free. Every feature is included, with no account and nothing to buy.
  • ImprovedThere is no daily word limit for dictation or file transcription.
  • ImprovedConverse is included for everyone, and coding prompt shortcuts no longer have a daily allowance.
  • NewDictation enhancement, Correction Memory, Smart Context, and Action Engine are unfinished, so they stay off until you turn on Beta features in Settings. No license key or tester approval is needed.
  • NewRead Aloud has a reading speed setting from 0.75× to 2×. It applies from your next reading.
  • ImprovedFile transcription lets you set the speaker count from 1 to 10, and Auto finds up to six voices. On Auto, laughter, crosstalk, and short fragments are less likely to be counted as extra speakers.
  • FixedA dictation is no longer dropped as silent when the microphone sends a few seconds of exact silence first, as a noise-gated microphone can. When the level meter reads nothing, the app now checks the recording itself before discarding it.
  • FixedThe first Read Aloud of a session starts about 0.1 seconds sooner again.
  • FixedLocal speaker labels keep short and repeated replies with the right speaker and are more reliable on recordings that are not in English.
  • FixedAn updated speaker labelling engine fixes a crash on some recordings.
  • FixedTimestamped and subtitle output keeps each word whole instead of splitting it across two cues.
  • FixedSpeech cleanup runs in its own process, so a crash there keeps your recording and no longer closes the app.
v0.19.1September 14, 2026

Steadier Downloads, Startup, and Windows Paste

  • FixedWindows paste preserves text already left on the clipboard when insertion cannot be confirmed, avoiding a second write while the target app may still be reading it.
  • FixedModel downloads keep their progress and cancellation controls when you leave a page and return. Download, deletion, and cancellation errors are visible.
  • FixedMinimized startup stays minimized, and duplicate automatic login launches no longer open the control panel.
  • FixedFailed History cleanup keeps the error and confirmation visible and preserves the previous retention setting.
  • ImprovedAgent Mode uses your normal dictation shortcut when enabled in settings. It starts disabled so existing settings cannot unexpectedly turn ordinary dictation into a submitted coding prompt.
  • ImprovedWindows paste diagnostics record bounded local timing and outcome metadata without dictated text.
v0.19.0September 9, 2026

More Reliable Long Dictations

  • FixedLong Whisper dictations preserve their endings more reliably, including speech after long pauses. Recording and History checks passed with Turbo and Large on clips up to ten minutes.
  • FixedLocal speech detection reduces invented closing phrases after breathing or silence while preserving quiet speech and intentionally spoken thank-you phrases.
  • FixedA slow Windows clipboard read can no longer block the automatic paste shortcut. Text stays on the clipboard when insertion cannot be confirmed.
  • ImprovedApproved testers get simpler dictation enhancement setup, refreshed model choices, and a test area for comparing cleanup results. Local and cloud providers use the same cleanup instructions.
  • ImprovedAI Enhancement brings coding prompt controls and dictation cleanup together, with separate settings for each.
  • ImprovedRead Aloud is included with Starter. Its compact player stays in place as word highlighting changes.
  • FixedRead Aloud recovers missing or shifted word highlights and skips code and other text that is not useful to speak.
v0.18.0September 6, 2026

Coding Prompts and Clearer Everyday Controls

  • NewAgent Mode turns a spoken draft into a clearer prompt through Claude Code, then pastes and submits it. It has its own hold key and daily allowance. Regular dictation pastes without submitting.
  • NewRead Aloud follows individual words in compatible Windows apps and updates their position more frequently while scrolling. This requires the new local timestamp-enabled voice model, about 326 MB. Existing voice choices are preserved.
  • NewRead Aloud playback shortcuts can be changed. New installs use Ctrl+Alt+R to start reading, with separate controls for pause, resume, and sentence navigation.
  • ImprovedConverse is included with Pro. Tester-only tools remain separate from paid Pro access.
  • ImprovedDictation has its own page. Settings and feature pages show essential controls first, with advanced options available on demand.
  • ImprovedThe overlay rests as a compact line, stays open during dragging, and returns to a visible display after monitor changes. Onboarding can continue while the GPU engine downloads.
  • FixedWindows paste no longer waits indefinitely for an app to expose its text. Transcripts remain available when delivery cannot be confirmed.
  • FixedWindows CPU engines include their required runtime libraries. File transcription, History feedback, copying, and microphone level recovery are more reliable.
v0.17.1September 5, 2026

Starter Is 1,000 Words a Day

  • ImprovedStarter now includes 1,000 transcribed words per day instead of 5,000. In five months nobody but the developer ever reached the old limit, so it never told anyone where Pro begins. Existing installs move to the new limit on their next day; today's count is kept.
v0.17.0September 4, 2026

Talk to Converse, and Read Aloud Marks the Page

  • NewConverse can be talked to instead of typed at. With a session running the microphone stays open, each spoken turn is found by its own pauses, and speaking over a reply cuts it off. It never opens the microphone without a running session, and it never answers a permission question for you
  • NewConverse gives spoken progress notes while it works, queues what you say while it is busy and answers it next, and speaks its permission questions so you hear it is waiting even when you are away from the window
  • NewRead Aloud marks the sentence it is on where the text lives, in the source app, instead of repeating it in the overlay. Works in Notepad, Chrome and Edge, and follows the page when you scroll
  • NewThe microphone picker sits beside the microphone permission, lists every input by name, says which one the next dictation will open, and has a Test button that listens for four seconds and tells you what it heard
  • NewThe app tells you when a dictation was not pasted. A paste that fell back to the clipboard, a paste that only made it to History, and a total loss each get their own message, the last one carrying the text itself
  • NewA message explains the slow first dictation after launch while the model loads, and a dictation that fails because no model is downloaded offers a button straight to the right Settings tab
  • NewGroq gains Whisper Large v3 beside the turbo model, and the OpenAI list says which transcription model is current, which is superseded, and what each costs per minute
  • ImprovedOne voice for the whole app. Read Aloud and Converse share the same voice setting, the voice list is grouped by accent with a preview on every row, and the letter grades are gone
  • ImprovedThe overlay rows are text-only capsules that float above the dictation button, with a live dot and a thin progress line instead of icons, dividers and counts
  • ImprovedSpeech synthesis runs in its own low-priority process, so Read Aloud and Converse no longer make the rest of the machine stutter while they speak
  • FixedThe dictation overlay opens when the app starts with your PC. Tray only and Minimized launches left you with the tray icon and no floating button until the first hotkey press
  • Fixed"Pause media while recording" really is off on Windows. The switch had been hidden but a saved preference could still pause and restart your video on every dictation
  • FixedThe Action Engine's Open Terminal starter template opens a terminal on Windows instead of failing silently
v0.16.1August 30, 2026

Silence Stays Silent

  • NewSettings search. One box that looks through every setting in the app and takes you to the row it found
  • ImprovedThe app no longer opens on a white window. Its fonts are bundled instead of fetched on every launch, and the window waits for real content before it shows
  • ImprovedLong recordings are cut at a pause instead of at a fixed 60 seconds, so a word is never split across the seam
  • FixedSilence is no longer transcribed. An open microphone with nobody talking used to come back as "Thank you" or a subtitle credit, and those are dropped instead of pasted
  • FixedYour system volume comes back after dictation. A failed restore used to bake the turned-down level in as your real volume
  • FixedCorrection Memory keeps Danish letters. Corrections containing æ, ø or å were silently dropped before
  • FixedOpening Settings no longer freezes the app for a second and a half
  • FixedReferral and creator discount codes work at checkout
v0.16.0August 27, 2026

Read Aloud and Converse

  • NewRead Aloud. Select text anywhere, press the hotkey, and PrivateTranscribe reads it back to you in a voice that never leaves your machine. It has its own page, 28 voices you can listen to before you pick one, and a player in the overlay that shows the sentence it is on
  • NewConverse. Hold a spoken conversation with an agent that can do real work on your machine. Talking over it stops the speech straight away and tells the agent what it missed, and a conversation survives restarting the app
  • NewYou can steer a read while it happens, from the overlay or the keyboard. Ctrl+Alt+Space pauses and resumes, and Ctrl+Alt+left or right skip a sentence at a time
  • NewOther apps are turned down while Read Aloud speaks, then put back where you had them
  • ImprovedRead Aloud starts on the first clause instead of waiting for a whole sentence, and the engine warms up in advance, so the first press no longer stutters
  • ImprovedTranscribing no longer overwrites your clipboard unless you ask it to
  • ImprovedThe Transcribe page takes its language from the languages you already said you speak
  • ImprovedEvery overlay sits in one column under the dictation button rather than scattering across the screen
  • ImprovedRead Aloud skips text that is not English and tells you, instead of garbling it
  • ImprovedReading aloud no longer takes the whole machine with it. Its threadpool is capped so everything else stays responsive
  • FixedStart on boot works. The switch was stuck on, claiming the app would launch at login when nothing was actually registered, and every attempt to change it failed with a warning blaming security software. It now reads the real Windows startup entry, so it tells the truth and can be turned on and off
  • FixedFile transcription no longer crashes when you cancel it, and stops transcribing in the wrong language
  • FixedA compound hotkey no longer misses the press when both keys land at the same instant
  • FixedTapping the dictation button over and over no longer walks your system volume down for good, and volumes come back even when a read never reports finishing
  • FixedExported settings now carry your agent name and your custom prompt
v0.15.0August 23, 2026

Every Language You Speak

  • NewYou can now name every language you speak, instead of picking one. Whisper is held to that set, so a Danish sentence with English words in it stops being guessed as a different language halfway through
  • NewLocal transcription now lets you choose how many CPU cores it uses. It previously took four regardless of how many the machine had
  • NewPrices are shown and charged in your own currency where Stripe supports it
  • ImprovedThe speed test now says how long a one-minute dictation actually takes, rather than only a multiplier
  • ImprovedThe custom dictionary is one flat list of terms. The three modes it had before did the same job three ways
  • ImprovedOptional analytics now also cover transcription speed, the compute mode, and the configured language and model. Nobody is opted in by this, everyone is asked again, and a previous no stays no. No audio, transcript text, filenames, window titles or API keys are ever sent
  • ImprovedEvery locked beta feature says so once, on the control that is locked, and links to how to get access
  • ImprovedThe update notice moved to the sidebar footer, right above the version marker, so the build you are running and the one waiting for you sit together
  • FixedCPU mode now really does keep the GPU out of it
  • FixedThe spoken-language picker no longer clips its list, hides its control, or leaves you unable to remove the last language
  • FixedYour first language is no longer forced onto every later dictation
  • FixedHistory entries stop being cut mid-word long before the preview is full
  • FixedStart on boot now refuses, and says why, when the app is running from a temporary folder that Windows will clean up. Before this it could leave a startup entry that showed an Electron error on every boot
v0.14.2August 15, 2026

Smaller Updates

  • ImprovedUpdates now download only the parts that changed instead of the whole installer
  • FixedA CUDA engine download that is already running is shown as running, instead of offering to start it again
  • ImprovedThe app now says what happens to a CUDA download if you quit part way through it
v0.14.1August 15, 2026

License Recovery and Clearer GPU Setup

  • NewYou can recover a lost license key yourself at privatetranscribe.com/license
  • ImprovedVoice call mute now holds your mute key down for the whole dictation instead of pressing it over and over
  • ImprovedWhisper models in the picker are sorted by size instead of internal order
  • ImprovedA refunded license now loses Pro, as the refund policy always said it would
  • FixedThe CUDA engine download showed the wrong size
  • FixedWhen Windows blocks the GPU engine the app says so, instead of blaming a driver update
v0.14.0August 13, 2026

Signed Installers, Voice Call Mute and Whole Long Dictations

  • NewWindows installers and the app itself are now code signed, so Windows no longer says the publisher is unknown
  • NewAdded voice call mute, which holds your voice app’s push to mute key while you dictate so nobody on the call hears you. A keyboard key or a mouse button both work
  • NewAdded a language step to onboarding, so the app knows what you speak before it recommends a model
  • ImprovedThe spoken language is detected once at the start of a recording and kept for the rest of it, instead of being guessed again on every segment
  • ImprovedThe AI now answers in the language you spoke
  • ImprovedAccuracy figures in the model picker follow the language you actually dictate in
  • ImprovedRaw dictation audio is no longer kept on disk after transcription
  • ImprovedLong dictations now split at a natural pause instead of on a timer
  • ImprovedCorrection memory asks for approval before it changes anything
  • FixedLong dictations no longer lose sections, and one failed chunk no longer takes the rest of the recording with it
  • FixedStopped the last half second of speech being cut off
  • FixedPasting on Windows now confirms the text actually landed before reporting success, and keeps the clipboard fallback when it cannot
  • FixedFixed tap and hold gestures being mistaken for each other
v0.13.11August 9, 2026

Pro Access in Packaged Builds

  • FixedAn active paid Pro license now bypasses the Starter daily word limit in packaged production builds
  • FixedUnsupported unpacked builds now link to the official Windows installer instead of showing a raw updater error
v0.13.10August 8, 2026

Immediate Pro Unlock and More Reliable Dictation

  • FixedA newly activated Pro license now removes the Starter word limit immediately, without restarting the app
  • ImprovedStable paid Pro access is now clearly separated from unfinished Beta features for approved testers
  • FixedLong dictations preserve complete audio more reliably and avoid truncated or hallucinated endings
  • FixedNavigation between the dictation overlay and control panel works reliably again
  • FixedTransparent overlay areas remain click-through instead of blocking the app underneath
  • FixedRepeated toast notifications are now deduplicated
v0.13.9August 7, 2026

Reliable Pasting Into Terminals

  • NewAdded OpenAI GPT Transcribe as a recommended cloud transcription model
  • FixedDictation now pastes correctly into terminal windows such as Claude Code, which need a different paste shortcut than ordinary apps
  • FixedPasting no longer fails when a dictation hotkey key is still held down as the text arrives
  • FixedRemoved the PowerShell step behind automatic pasting, which some antivirus tools flagged as suspicious
  • FixedYour clipboard is now restored only after the paste has finished, so nothing gets overwritten mid-paste
  • FixedThe Transcribe language menu no longer gets cut off near the bottom of the window
  • FixedFixed the misspelled PrivateTranscribe entry in the Windows Start Menu
  • ImprovedHardened app navigation, model downloads, release credentials, and local API key storage
v0.13.7July 27, 2026

Hotkey Control, Analytics Choice and Startup Fix

  • NewAdded a tray toggle that pauses the dictation hotkey for the current session
  • NewAdded a Settings control for changing optional product analytics consent at any time
  • ImprovedLimited product analytics to known events, removed duplicate tracking, and set a 12-month retention period
  • FixedPrivateTranscribe now respects when its startup entry is disabled in Windows, and Settings reflects the real Windows state
v0.13.6July 13, 2026

Clearer Starter Reset Timing

  • ImprovedThe Starter usage card now shows a live hours-and-minutes countdown until the daily allowance resets
v0.13.5July 13, 2026

Faster Local Dictation & Usage Clarity

  • NewAdded a Dashboard card showing Starter words used, words remaining, and the daily reset time
  • ImprovedLocal transcription now starts loading while you speak, reducing the wait after recording on cold starts
  • FixedImproved overlay hide and show behavior across taskbar snapping, tray actions, and Settings
  • FixedAllowed slower local model cold starts enough time to finish without leaving failed attempts hanging
v0.13.4July 10, 2026

Responsiveness Improvements

  • ImprovedCUDA engine setup, model cache migration, and hardware detection now run without blocking app startup
  • ImprovedParakeet model deletion no longer blocks the main process
  • ImprovedAudio transcription file handling now uses async file I/O for smoother processing
v0.13.3July 9, 2026

Production Settings Cleanup

  • FixedProduction Settings now labels the storage tab as Data & Storage instead of Developer
  • FixedRemoved the empty Diagnostics & Data header when developer tools are hidden in production
v0.13.2July 8, 2026

Entitlement Gating & Brand Fixes

  • FixedFixed packaged builds accidentally treating themselves like development builds, which unlocked gated paid and tester-beta features without a valid entitlement
  • FixedRestored Pro surfaces in Settings to the website purple accent
v0.13.1July 5, 2026

CUDA Engine & Starter Limit

  • ImprovedPublished the CUDA engine update path for the latest GPU package
  • NewStarter now includes 5,000 private words per day with all Whisper models and CUDA acceleration available
  • FixedExisting Starter installs are migrated from the old daily cap to 5,000 words
v0.13.0June 24, 2026

Tester Approval Flow & Email Overhaul

  • NewManual tester approval flow with Discord Approve and Deny buttons
  • NewName and use-case fields on the tester signup form
  • NewLicense key auto-emailed to approved testers
  • ImprovedEmail templates aligned with brand writing style
  • FixedDiscord webhook buttons fixed with component support
  • FixedEmail send failures now surfaced instead of swallowed
  • FixedTaskbar snap defaults to on
v0.12.2June 23, 2026

Feedback Screenshots & Discord

  • NewAdded screenshot attachments to in-app feedback - paste (Ctrl+V) or add images directly in the feedback dialog
  • NewFeedback now forwards to Discord with screenshot links and image previews so reports arrive in one place
  • ImprovedDifferentiated Discord feedback category colors so feature ideas and general feedback are easier to tell apart
  • ImprovedFeedback now always includes app version and system info with a clear privacy notice instead of an opt-in checkbox
v0.12.1June 22, 2026

Windows Uninstall Cleanup

  • FixedImproved Windows uninstall cleanup so a manual uninstall can remove app data and model caches without affecting updater installs
  • FixedPrevented uninstall cleanup prompts from appearing during app updates or silent uninstalls
v0.12.0June 22, 2026

Dictionary Modes & Correction Memory Beta

  • NewAdded dictionary modes so words can be saved as hints, exact spelling repairs, or higher-priority terms
  • ImprovedMerged the approved-tester Correction Memory beta into the dictionary page so vocabulary and learned fixes live together
  • ImprovedKept the Correction Memory beta focused on useful word-level repairs
  • FixedEnforced production entitlements so development previews cannot unlock paid or tester-beta features
  • FixedValidated Whisper model files before use so broken or partial downloads are not treated as ready
v0.11.4June 17, 2026

Installer Build Fix

  • FixedPrepared a new Windows build version after the unsigned local installer build fix
v0.11.3June 16, 2026

Local Installer Builds

  • FixedFixed local Windows installer builds without code-signing privileges while preserving app executable metadata
v0.11.2June 16, 2026

Windows Build Compatibility

  • FixedFixed Windows installer builds with electron-builder 26.8.1 by removing an unsupported signing option
v0.11.1June 15, 2026

Startup & Installer Polish

  • NewAdded the startup-at-login choice to onboarding so new users can decide during first setup
  • FixedRestored Windows executable resource editing so packaged installer builds embed the PrivateTranscribe icon while still skipping code signing
v0.11.0June 14, 2026

Long Dictation & Setup Clarity

  • NewAdded support for long-running dictation sessions with stronger recorder handling for longer speech input
  • ImprovedImproved dashboard layout, setup summary, stat pills, and microphone configuration presentation
  • ImprovedSimplified tester feedback submission and aligned feedback categories with the backend
  • FixedHardened dictation finalization to reduce race conditions and avoid losing the tail end of longer dictations
  • FixedRestored the overlay microphone visualizer after system wake and resume scenarios
v0.10.5June 11, 2026

Tester Feedback & Setup Polish

  • NewAdded in-app structured feedback collection so testers can report issues directly from the UI
  • ImprovedGated full dictation setup on real Whisper model download so installs are more reliable on first launch
  • FixedDeferred first-open diagnostics to prevent setup lockups on slower machines
  • ImprovedRecommended large-v3-turbo on high-VRAM CUDA setups for the best out-of-the-box accuracy
  • FixedClarified transcription output language labels and prevented stale translation toggles from affecting results
  • FixedFixed incorrect auto-start registration so PrivateTranscribe launches cleanly on boot
  • ImprovedAdded opt-in text repair diagnostics to help diagnose split-word spacing issues
v0.10.4June 4, 2026

Early Access Feedback System

  • NewBuilt in-app feedback panel so early-access testers can report bugs and suggest improvements without leaving the app
  • NewAdded Supabase-backed feedback collection Edge Function with health-check endpoint
  • ImprovedAdded CI workflow for feedback backend deployment alongside the main app
v0.10.3June 3, 2026

Transcription Repair & Support Tools

  • FixedRepaired split transcription spacing so dictated text no longer arrives with missing mid-word spaces
  • ImprovedFixed support email templates so users can describe issues more clearly when reporting problems
  • FixedHardened support feedback tooling for more reliable issue reporting
v0.10.2May 31, 2026

Overlay Taskbar Snap & Diarization Stability

  • NewAdded taskbar snap mode - overlay can now dock to the Windows taskbar for a cleaner desktop dictation experience
  • FixedOverlay recovers properly after display sleep, wake, and z-order changes - no more ghost overlays
  • FixedShell-level diarization (speaker detection) is more stable with safer thresholds, max speaker caps, and no crash-prone clustering
  • ImprovedStopped repeated Whisper model reloads between dictation sessions for faster recording start
  • ImprovedParallelized local Whisper preparation steps to reduce recording latency
  • FixedFixed partial legacy model cache migration that could leave the wrong model selected after upgrade
  • NewAdded combined tap-and-hold dictation mode - tap to toggle, hold for push-to-talk
  • ImprovedNormalized transcript punctuation spacing so apostrophes and commas land reliably
v0.10.1May 12, 2026

File Transcription v2

  • NewAdded production-grade Transcribe File workflow for local audio and video files
  • NewAdded automatic speaker detection with clear Speaker 1 / Speaker 2 style labels
  • NewAdded timestamped output and SRT subtitle export for easier reuse in editing and publishing workflows
  • ImprovedAdded optional noise reduction and persisted file transcription settings
  • FixedPreserved TinyDiarize speaker-turn markers from realtime whisper.cpp output for more reliable speaker formatting
  • FixedFixed a startup crash when reading speaker detection settings before initialization
v0.8.4May 6, 2026

Windows Hotkeys & Overlay Stability

  • FixedSafer Windows hotkey routing - standard shortcuts like Ctrl+Space stay on Electron global shortcuts, while native-only keys use the Windows listener
  • FixedMouse4 / Mouse5, modifier-only combinations, and Danish/OEM keys such as ½ now register reliably on Windows
  • ImprovedPreserves the user-visible hotkey label while normalizing native Windows key codes behind the scenes
  • ImprovedTray wording now consistently refers to the floating UI as the Overlay
  • FixedResolved dependency audit issues and updated native dependencies for the newer Electron runtime
v0.8.3May 1, 2026

CUDA Engine Reliability

  • FixedImproved CUDA package validation for split runtime builds so GPU acceleration is detected more accurately
  • ImprovedAdded clearer CUDA engine status updates in Settings → General
  • FixedLanguage auto-detection is now passed explicitly to Whisper for more predictable transcription behavior
  • ImprovedAdded long-file transcription coverage to protect larger local audio workflows
v0.8.2April 30, 2026

Packaged Whisper Runtime

  • ImprovedPackaged required Whisper and CUDA runtime libraries with the Windows app so installs are more self-contained
  • ImprovedAdded startup diagnostics when CUDA cannot be used, making CPU fallback easier to understand and support
  • ImprovedDefault transcription language now starts on auto-detect for a smoother first-run experience
  • FixedClarified when the CUDA engine is active versus when transcription has fallen back to CPU
v0.8.1April 30, 2026

CUDA Status Clarity

  • ImprovedClarified CUDA active-state copy so users can see whether GPU acceleration is actually being used
  • FixedImproved CPU fallback messaging when CUDA is unavailable or inactive
v0.8.0April 30, 2026

Long Audio & Engine Status

  • NewLocal file transcription now handles long audio files more reliably
  • ImprovedEngine status is shown even when the Whisper server is idle, so users can see what backend is actually ready
  • FixedReduced Whisper server restart churn between requests
  • FixedFixed CPU-only mode enforcement and CUDA-to-CPU fallback handling
v0.7.0April 29, 2026

Whisper Engine Foundation

  • NewAdded Windows CPU whisper-server packaging so local transcription works even without CUDA
  • ImprovedImproved CPU/GPU engine labels and fallback indicators so performance state is clearer
  • FixedAdded binary size guards to the CUDA workflow to catch broken runtime uploads earlier
  • FixedHardened startup behavior when CUDA binaries or drivers are unavailable
v0.5.2April 24, 2026

Update Install Stability

  • FixedFixed an update-install crash by making the app quit path safer during auto-updates
  • ImprovedSkipped model cache prompts during automatic updates so installs can complete without unnecessary interruption
v0.5.1April 24, 2026

Release Polish

  • FixedRemoved a manual before-quit event path that could interfere with update installation
  • ImprovedCleaned up lint, formatting, and release-test expectations after the v0.5.0 CUDA work
v0.5.0April 23, 2026

CUDA Runtime & Onboarding Upgrade

  • NewAdded CUDA binary version tracking with SHA256 hashes and stale-runtime detection
  • NewAdded automatic CUDA binary updates after app updates to keep the GPU runtime aligned
  • ImprovedAdded CUDA download to onboarding and gated setup on the current CUDA binary
  • ImprovedSimplified onboarding with clearer hardware selection, less jargon, and a more realistic microphone test
  • ImprovedRemoved Parakeet from the recommended setup path and cleaned up stale Parakeet settings
  • FixedFixed CUDA download safety issues including concurrent downloads, stale banners, and safe failure handling
  • FixedFixed overlay wake/reconnect positioning and several reviewed security/stability issues
v0.4.1March 29, 2026

Overlay & Action Engine Beta Reliability

  • FixedFixed overlay context menu overflowing screen edges - menu now flips direction based on available space instead of always opening upward
  • FixedFixed approved-tester Action Engine beta commands firing twice on overlapping transcript segments - added per-action debounce (2 s default)
  • ImprovedHistory empty state now shows your actual configured hotkey so new users know how to start
v0.4.0March 23, 2026

Performance & Reliability

  • ImprovedEliminated 3-4s recording start delay by caching microphone device enumeration
  • ImprovedPooled microphone stream between recordings to remove getUserMedia latency on every dictation
  • ImprovedUpgraded overlay to screen-saver window level on Windows for guaranteed always-on-top behaviour
  • FixedOverlay no longer intercepts mouse clicks in transparent areas
  • FixedFaster on-top self-repair loop so overlay recovers if another window pushes it down
  • ImprovedTray tooltip simplified to app name for cleaner taskbar presentation
v0.3.4March 21, 2026

Pre-Launch Polish

  • ImprovedRewrote gated-feature locked-state copy to be clearer
  • ImprovedAudited paid and tester-beta entitlement enforcement across gated surfaces
  • FixedAdded root-level ErrorBoundary for graceful crash recovery
  • FixedNormalized support email handling across the app
  • FixedFixed dictionary word replacement and streak date calculations
v0.3.0March 15, 2026

Action Engine & Smart Context Beta

  • NewAdded the approved-tester Action Engine beta for voice-triggered commands and workflows
  • NewAdded the approved-tester Smart Context beta for IDE-aware dictation
  • NewAdded the approved-tester Correction Memory beta for learning from edits
  • ImprovedAuto-disable translation for unsupported Whisper models
  • ImprovedWarning when Turbo model is used with translation enabled
v0.2.0March 1, 2026

Settings & Onboarding Overhaul

  • NewRedesigned settings with proper navigation sections (dictionary, AI models, agent config, prompts)
  • NewGPU status card with driver download links and re-detection
  • NewCPU vs GPU transcription benchmark with honest performance labeling
  • ImprovedImproved onboarding completion flow with clearer copy
  • FixedHardened Linux CPU-only hardware detection
  • FixedFixed overlay cancel button only showing for mouse-click activations
  • FixedFull app restart on "Reset app data" instead of renderer-only reload
v0.1.5February 15, 2026

UI Polish & Stability

  • ImprovedStandardized dropdown and select styling with consistent surface tokens
  • ImprovedDashboard empty state shows hotkey CTA matching onboarding
  • FixedFixed vendor names, VRAM display, and CUDA detection copy
  • FixedFixed local date handling for streak calculations
  • FixedStabilized Windows push-to-talk and drag behavior
v0.1.0February 1, 2026

Initial Release

  • Launch100% local speech-to-text powered by OpenAI Whisper
  • NewGlobal hotkey activation from any application
  • NewAuto-paste transcription directly to your cursor
  • NewMultiple Whisper model sizes (tiny → large-v3-turbo)
  • NewVoice-reactive overlay with real-time feedback
  • NewNative Windows desktop app with system tray support
  • NewAdded the approved-tester AI Enhancement beta for grammar and formatting experiments
  • NewFile upload with multi-format audio support

PrivateTranscribe is in active Windows beta. Release notes focus on user-facing changes, stability improvements, and important platform fixes.