Changelog

# OpenWhispr 1.10.3 The assistant can now act in other apps: it drafts and sends email, posts to Slack, and works with Linear and GitHub issues, always behind a card you approve. Notes get templates for the AI Summary, actions that answer in a docked chat, and real tables. OpenWhispr also uses far less memory when idle, shows a plain-language error when your API key or provider fails, and fixes a long list of bugs on Linux, Windows and macOS.

🌟 Highlights

• 🔌 Connectors — The assistant can draft and send email (Gmail), post to Slack, and work with Linear and GitHub issues. Nothing goes out until you press Send on a card showing exactly what will be sent. Paid plans; connect in Settings → Integrations. (#2331, #2369, #2385, #2401, #2402)

• 📝 Note templates and actions — Templates shape the AI Summary, and actions like Follow-up email answer in a chat docked beside the note instead of overwriting your summary. (#2397, #2557, #2578, #2598)

• 📊 Tables in notes — Tables stay tables when you edit a note, and a formatting toolbar appears when you select text. (#2255)

• 🪶 Lighter when idle — Search and meeting detection only run when needed, cutting idle memory use from about 500 MB. (#2299, #2371)

• 💬 Clear error messages — When your own API key or provider fails, you get a plain-language message with a link to fix it. (#2510)

✨ Also new

Echo cancellation for meetings, an option to keep the local model loaded, a folder picker in the meeting notification, dictation audio that survives a crash, a macOS permission helper in onboarding, Apple sign-in on Windows and Linux, and live speaker names on Linux.

🛠 Fixes

• Windows: Teams meetings record everyone for the whole call again. (#2233)

• Transcription: API key changes apply without a restart, and AssemblyAI, Corti and Gemini Live no longer drop words.

• Assistant: Answers paste as plain text, follow-ups don't repeat actions already sent, and images in replies are no longer loaded automatically.

• Linux: GPU rendering is back (except on NVIDIA), Settings no longer lags, and paste waits for you to release held keys.

• Notes and meetings: Assistant edits to a note can be undone and never wipe fields by accident, clearing a note now syncs, and fixes for Join & transcribe, live transcripts on reload, and TXT export. (#2602, #2604)

⚠️ Heads-up

The Windows portable .exe is discontinued. Use the installer instead. (#2301) The full list is in the changelog.

🙏 Thanks

Thanks to @Chadpiha, @boseq, @greatcoat, @hsusul, @IdrisGit, @clarkmaan, @Nathan-Roll1, @gabrielste1n, @sochotnicky, @kuleje, @TrackerGit, @zhongwater123, @emanuelet, @marselsel, @folks-samuel-cb, @STiFLeR7, @Adityakk9031, @stantheman0128, @cadufpolis, @Abhishek-B-R, @dajiaohuang and @hiro-nikaitou for their work on this release. 🎉 And thanks to @jaszczurovsky, @TaiFeng, @jsholcomb, @wadamek65, @KishenG and @amarpreet2209 for reporting and narrowing down the Teams capture bug (#1265).

Prebuilt Windows system audio helper binary for meeting transcription.Captures system audio via WASAPI process loopback (exclude mode), hearing every application on every output device while excluding OpenWhispr's own audio. Requires Windows 10 2004 or later; older systems automatically fall back to Chromium display-media loopback.

# OpenWhispr 1.10.2 A hotfix for 1.10.1. Generate AI Summary with your own API key or an enterprise provider no longer times out on a real transcript, a recording with nothing substantive in it gets a one-line summary instead of a blank page and an error, and OpenRouter models can read screenshots again.

🛠 Fixes

• ⏱ Generate AI Summary no longer times out with your own key — Every bring-your-own-key request shared the 30-second deadline meant for dictation cleanup (enterprise providers had 60 seconds), and a timeout was retried three more times as if the connection had dropped. A long transcript through a reasoning model such as GPT-5.6 Terra spun for about a minute and a half, billed four requests, and ended with a timeout error and no note. Note formatting now waits up to ten minutes, matching OpenAI's own client default, and a request that does hit its deadline fails once instead of being re-sent. Dictation cleanup keeps its 30-second deadline everywhere. (#2204)

• 📝 A recording with nothing in it gets a one-line summary instead of an error — Greetings, filler or a sound check could end Generate AI Summary with an enhancement error, because Detailed Notes stripped the empty content, the model omitted every section, and the blank-response safeguard from 1.10.1 correctly refused to save the result. Generate Notes and Detailed Notes now answer with one factual sentence in your output language. Untouched default prompts move to the new text automatically; an edited prompt is left alone. (#2201)

• 👁 OpenRouter models can read screenshots again — OpenRouter returns model IDs with a vendor prefix such as google/gemini-3.5-flash-lite, while the model registry stores the upstream ID, so the vision-capability lookup found nothing and screenshot commands were rejected as unsupported. The lookup now retries without the prefix for OpenRouter only; other providers are unchanged. (#2203)

The full list is in the changelog.

🙏 Thanks

Thanks to @Chadpiha and @fthozdemir for their work on this release. 🎉

# OpenWhispr 1.10.1 Version 1.10.1 brings a fresh round of improvements and fixes to the desktop app. Local dictation gets noticeably faster, team spaces get a single roster and a single settings dialog, and a long list of transcription and dictation bugs are closed.

🌟 Highlights

• ⚡ Faster local dictation — Parakeet and Orukeet decode their 15-second segments concurrently, the app hands local engines a raw 16 kHz copy instead of making FFmpeg unpack WebM/Opus first, and on Windows the microphone is resolved once at launch rather than on your first hotkey press. (#2178, #2179)

• 🔑 Bring-your-own-key Deepgram and AssemblyAI work again — Deepgram rejected every connection with a 401 because a raw API key must be presented under its Token scheme, not Bearer; AssemblyAI died seconds in because Note Recording's 24 kHz frames were measured against dictation's 16 kHz. (#2149)

• 👥 One roster, one dialog, per space — A space's menu opens a single Settings dialog, and Members is now one flat list of everyone with access, whether added directly or through a group, with rows marked _via Design_ where a group grants it. Creating a space no longer requires a group. Teams are called Groups in workspace copy across all 11 locales. (#2186, #2194)

• 🎙 Windows pins the real microphone again — Chromium lists the Windows default input twice, once under its communications alias, and the selection code fell back to a generic default it never cached, re-resolving the device on every press. (#2180)

• 🧠 Local models size their context to the request — Generate Notes on a meeting over roughly 80 minutes failed with a raw llama.cpp error. (#2155)

✨ New

• New note is a split button — + New note creates a note and starts recording, from any view; its chevron adds Assistant chat and Team space. The tray menu leads with Start listening / Stop listening, Ask assistant and Start meeting recording, and the dictation pill picks up the same quick actions. (#2163, #2199)

• Notes open on their AI Summary — A note that already has a summary opens on it instead of the raw tab, and the "Enhanced" label is gone: the tab, the toggle and every mention now read AI Summary, matching the Generate AI Summary action that produces it, in all 11 locales. The raw tab becomes Your notes. (#2163, #2196)

• Invite your team from the sidebar — Workspace owners and admins get an Invite your team row that opens the existing invitation dialog. (#2181)

• A built-in emoji picker — Searchable, categorised, keyboard-navigable, with recents, translated into ten languages. It replaces the OS panel, which behaved differently everywhere and never worked on Linux. (#2186)

• Upload takes video and almost any audio container — MPEG audio, .mp4, .mov, .mkv and .3gp with a speech track are accepted, and a rejected drop now says why. (#2193)

• Detailed Notes gets a shorter, sharper prompt — Rewritten and tested side by side against the old one on real 30-, 60- and 90-minute meetings. Action items now use the same - [ ] Action — Owner shape the editor's owner tagging can actually match. An edited prompt is left alone.

🛠 Fixes

• Transcription — whisper.cpp no longer drops words at its 30-second window boundaries, uploads and meetings use the local engine you chose in onboarding instead of falling back to Whisper, and the download tray says Installing once extraction starts. (#2152, #2033, #2137)

• Dictation — A cleanup reply the model cut short is no longer pasted as though it were complete, on Gemini either; the pill stops restacking on every live-transcript chunk and stays hoverable on Linux; a snippet trigger starting with the assistant's name expands instead of waking it; and a dictation started from the Windows tray pastes where you were working. (#2130, #2092, #2185, #2161, #2153)

• Meetings — The recording timer survives switching notes, a meeting that ends on its own still offers Generate AI Summary, the post-meeting restart card is gone, and Windows system-audio capture ignores a stopped helper's last words. (#2162, #2192, #2197, #2017)

• Assistant — A "Separate vision model" override switched on but never configured no longer routes screenshot commands to a keyless provider, tables render instead of collapsing into literal pipes, and long answers stream without re-parsing the whole reply per token. (#2139, #2131, #2171)

• Notes — Generate AI Summary no longer saves your raw transcript back as the summary when the model returns nothing, and date group headers no longer appear twice. (#2177, #2000)

• AI models — Learned corrections skip everyday words, so fixing "why" to "what" once no longer teaches a rule that rewrites later dictations, and OpenCode Go works as a custom OpenAI-compatible provider. (#2178, #2146)

• Settings & CLI — Reset app data restarts the app so history keeps working, and the local CLI bridge rejects malformed limit and folder_id values instead of coercing them. (#2134, #2015)

The full list, including every smaller fix, is in the changelog.

🙏 Thanks

Thanks to @Chadpiha, @boseq, @greatcoat, @hsusul, @IdrisGit, @Nathan-Roll1 and @L4XB for their work on this release. 🎉

# OpenWhispr 1.10.0 The biggest release since 1.9. The desktop app has a new look, notes take care of themselves from the moment you create one, local transcription gets a new recommended model that runs up to three times faster on Apple silicon, and enterprise workspaces can hand every member a managed speech-to-text deployment with no keys on their machines.

🌟 Highlights

• 🎨 A new shell — The main window's content sits in an inset, rounded container with a top bar: sidebar toggle, page title, and ⌘K search across notes and transcripts. Home groups transcriptions by day, shows the week's meetings as day cards with a hover-revealed Join & transcribe, and every empty state now says what to do next.

• 📝 Notes that run themselves — A new note starts recording the moment you create it. When the recording ends, the Transcript tab offers Generate AI Summary, which produces accurate, comprehensive meeting notes with decisions, owners, and due dates. Draft a follow-up email sits in the ask bar as the default action.

• 🗣 Orukeet, the new recommended local model — A multilingual Parakeet fine-tune from Oruk that runs on the sherpa-onnx runtime you already have. It is the first tab in the local model picker, and a fresh install choosing the local route lands on it. (#2085)

• ⚡ Local models up to three times faster on Apple silicon — sherpa-onnx now ships the arm64-native ONNX Runtime; an 11-second clip drops from 605 ms to 173 ms on Orukeet with identical transcripts. (#2116)

• 🏢 Managed enterprise speech-to-text — Workspaces on Azure OpenAI or AI Foundry can distribute a transcription deployment to every member, with no API keys on user machines. Members need 1.10.0 or later. (#1964)

• 🔄 Automatic updates — Updates download in the background and install when the app next quits. On by default for new installs; existing installs opt in under Settings → System. The update-available popup is gone. (#2120)

• 🌍 Arabic — The eleventh UI language, with full right-to-left layout. (#1869)

✨ New

• Detailed Notes and Follow-up email — Two built-in actions join Generate Notes. Detailed Notes keeps every client, project, number, and date from the call, separates decisions from proposals, and gives action items an owner and a due date. Follow-up email drafts the email you would send afterwards. Every built-in prompt stays editable, and an edited prompt is never overwritten by an update.

• Redesigned note editor — A larger title, combined date and attendees capsule, one toolbar row with Transcript, Notes, and AI Summary on the left and the record control plus a split Share on the right. Recording widens the record circle into a stop, waveform, and timer pill. Export lives in the Share dialog. The Transcript tab is always reachable, with an empty state that offers to start recording.

• The assistant edits your dictionary and snippets — Three new tools fetch a snippet by trigger, add or remove dictionary words, and create, replace, or delete snippets. The CLI bridge gains matching routes. (#2119)

• Transcribe files from the terminal — openwhispr transcribe runs whichever local model the app is set to use. Audio never crosses the bridge. (#2121)

• Streaming from Deepgram, AssemblyAI, and Gemini Live — All three are bring-your-own-key streaming providers for dictation. (#2062, #2063)

• Cohere Transcribe runs locally — The most accurate offline model OpenWhispr ships, 5.42% average WER across 14 languages, on the same sherpa-onnx server as Parakeet. (#1949)

• gpt-transcribe is the new OpenAI default — OpenAI is retiring whisper-1 and the gpt-4o transcribe models; nothing stored is rewritten and the old models remain selectable. (#2076)

• Insights and an opt-in leaderboard — A usage view computed locally, backfilled from your existing history, with an optional leaderboard for colleagues who also join. (#1959, #2011, #2075)

• Astra and Fable 5.1 — Available as reasoning models. (#2107)

🛠 Fixes

• Meetings — A renderer reload mid-recording left the main process holding the session, so every later start failed with "Operation in progress" until the app restarted. Sessions now stop when their window navigates away, and a restart from the same window retires the stale one. Recordings that went silent partway through on macOS recover the system-audio tap, and playback no longer raises a meeting prompt. (#2040, #2071)

• Dictation — A push-to-talk that hits the safety ceiling keeps its transcript on the clipboard instead of losing it. A failed paste no longer costs you the dictation. Windows and Linux media pause/resume no longer block the main process. (#2102, #1979, #2094)

• Transcription — Long custom dictionaries are no longer cut in half. Local dictations no longer come back as pieces of your own dictionary. Cohere Transcribe accepts underscore and padded locale codes. (#1960, #1929, #1967)

• Routing — Signing in no longer silently reroutes dictation to OpenWhispr Cloud, and note recordings no longer go to the cloud for a user who is Local everywhere else. (#2093, #2097)

• Windows — Local transcription works without the VC++ redistributable, Parakeet starts on machines with an older system ONNX Runtime, and winget manifests are published from CI with permanent URLs. (#2074, #2089)

• macOS — Accessibility features wait for their permission instead of prompting during first launch. Plus-addressed email accounts can sign in. (#2101, #2088)

• Linux — Hyprland punctuation hotkeys validate, and auto-paste no longer drops keystrokes on GNOME Wayland. (#2078)

• Notes — Notes created offline reach mobile promptly, Granola imports keep notes that share a title and date, and the transcript follows the live meeting again. (#2065, #1955, #1896)

• Security — Two high-severity advisories in bundled dependencies were closed. (#2079, #1986)

The full list, including every smaller fix, is in the changelog.

🙏 Thanks

Thanks to @Chadpiha, @boseq, @greatcoat, @hsusul, @IdrisGit, @Nathan-Roll1, @Zeljaa, @alekc, @dajiaohuang, @xlurie, @AdityaPainuli, @B0-KA and @clarkmaan for their work on this release. 🎉

# OpenWhispr 1.9.2 A repair release for 1.9.1. If you are on Windows, this one matters — sign-in was completely broken.

🌟 Highlights

• 🔑 Windows sign-in works again — Every provider button on the Windows desktop app had gone dead since 1.9.1. Connecting a calendar and Stripe checkout broke the same way.

• 📝 Transcription works in packaged builds again — xAI dictation, re-transcribing a saved recording, and bring-your-own-key file uploads all failed on Windows, macOS, and Linux. Only development builds were unaffected, which is how it reached release.

• 🔊 Windows meetings capture everyone again — Some machines recorded your microphone and nobody else, with no error anywhere in the app.

• 🔔 Meeting prompts survive the first call — Closing a prompt any way other than its own buttons stopped every later meeting from ever prompting again.

🛠 Fixes

• Windows sign-in — Browser links were routed through explorer.exe, which silently opens a File Explorer window instead of the browser for any URL carrying a query string — and every sign-in URL carries one. Those links now open directly, while the ones explorer.exe can deliver still go through it, so a browser launched from a meeting link keeps being captured. (#1932)

• Transcription — A model-registry import added in 1.9.1 was missing the attribute Node's module loader requires for JSON, so three transcription paths threw needs an import attribute of "type: json" in packaged builds while ordinary dictation kept working. (#1908)

• Meetings — On Windows the system-audio helper could start up reporting itself healthy and then capture nothing but digital silence. It now compares its own output against what your speakers or headphones are really playing, and a recording that hits this switches to the working fallback a few seconds in and keeps the rest of the call. (#1935)

• Meetings — A prompt closed by a compositor, a failed load, or onboarding taking the screen left its detection marked as still on screen, so nothing prompted again for the rest of the session. On Linux the prompt also fired on OpenWhispr's own microphone use, and now waits until the audio server has said which application owns the capture. A second call raises a fresh prompt once the first one's microphone has gone quiet, while an ongoing call still prompts only once. (#1918)

• Meetings — Meeting cards dismiss with an 80px horizontal swipe, like desktop notifications. Detection prompts, calendar reminders, and the auto-end restart offer all take the same path as their close button. (#1906)

🙏 Thanks

Thanks to @greatcoat, @Chadpiha and @IdrisGit for their work on this release — @IdrisGit both reported and fixed the meeting-prompt bug — and to @VictorZhang01 and @KishenG for the reports behind two more of these fixes. 🎉

Prebuilt Windows system audio helper binary for meeting transcription.Captures system audio via WASAPI process loopback (exclude mode), hearing every application on every output device while excluding OpenWhispr's own audio. Requires Windows 10 2004 or later; older systems automatically fall back to Chromium display-media loopback.

# OpenWhispr 1.9.1

🌟 Highlights

• ⏹ Meetings stop when the call stops — Auto-end drops the 60-second countdown, so recordings finish about a minute sooner. If it called it wrong, a 30-second card puts you back into the same note.

• 🎙 Gemini for transcription — gemini-3.5-transcribe joins OpenAI, Groq, and Mistral as a bring-your-own-key provider.

• 📥 Granola import — Bring your meeting notes across from Granola's CSV export with dates, transcripts, and speaker names intact.

• 🔒 Account deletion stays in its lane — Deleting one account no longer touches another account's notes, your workspace content, or your downloaded models.

✨ Also new

• A dictation pill for Agent mode, on the opposite edge of your display

• Screen Context permission during setup

• Your assistant can find free calendar slots, now 31 days out instead of 7

• Buy Enterprise in-app, with one-click entry to the web console

🎙 Dictation

• A cyan-to-indigo processing glow replaces the rainbow beam

• Assistant answers paste back into the field you were typing in, or fall back to your clipboard

• Recording starts 4–5 seconds sooner for local-mode and signed-out users

• No trailing space after Chinese or Japanese text

🛠 Fixes

• Performance — 1.9.0's recording-time CPU spike is fixed, and GPU packs that 1.8.3 deleted are now repaired at launch.

• Meetings — On Windows, meetings that recorded you but nobody else: a browser opened from the meeting link landed inside the process tree that system-audio capture excludes.

• Calendar — Recurring meetings on shared Microsoft calendars have names again.

• AI models — Stray tags stop leaking whatever their capitalization, and an Anthropic reply that hits its token limit no longer crashes cleanup.

• Elsewhere — Settings stays open when you dismiss a dropdown, debug logs record which release produced them, packaged builds block untrusted file navigation, and the CLI bridge handles UTF-8 correctly.

🙏 Thanks

Thanks to @hsusul, @AdityaPainuli, @Chadpiha, @boseq and @greatcoat, for their work on this release, and to @lingfish, @Nagell, @RonanCJ and @xlurie for the reports behind several of these fixes — and a warm welcome to first-time contributor @zhongwater123! 🎉

# OpenWhispr 1.9.0

🌟 Highlights

🗣 One Voice Assistant

The separate always-on-top chat window is gone. The dictation pill is now the single entry point: speak a command and the answer streams into a floating panel with the chat's full toolset — search and edit your notes, check your calendar, search the web, use your clipboard, and remember context when you're signed in. Highlighted text is still rewritten in place. Follow up by typing or by pressing the hotkey again to speak into the same conversation, and everything is saved and resumable. Your custom dictionary now reaches the assistant too, so replies come back using your names and jargon. The pill itself is redesigned: an idle orb that grows into a capsule with a live, level-driven waveform while you speak, a processing state that never jumps around mid-cycle, and cancel right where you're looking.

🚀 Guided setup, rebuilt

First run now walks you through sign-in, permissions, language, use cases, hotkey capture, and live dictation and assistant demos — then sets you up on OpenWhispr Cloud, your own API keys, a local model, or your organization's enterprise config, validating providers as it goes and downloading models in the background. Signing back in later uses the same compact screen instead of a form buried in a full-size window.

📝 Notes that know who said what

Meeting notes are generated with real speaker names instead of guesses: every transcript line carries its resolved speaker, your own track carries your name, and the prompt is forbidden from inventing identities — so notes stop assuming you're whoever got named in conversation. Action items generate with an owner attached, and typing @ in any note opens a picker of your attendees and known speakers that exports as a readable link. The notes surface also got a design pass — a voice-reactive recording pill, a translucent bottom bar your content scrolls under, and voice notes recorded straight into the chat input for review before you send.

⏹ Meetings that end themselves

Forget to stop a recording and OpenWhispr now notices the call is over — your meeting app released the microphone, both audio channels went quiet, or the app exited — and ends it after a 60-second countdown you can cancel with Keep recording. Remote voices still playing defer the countdown, so a quiet-but-live meeting is never cut short.

✨ Also new

• Timestamped transcripts and SRT export for uploads — Audio you upload or ingest from a URL, transcribed with your own OpenAI, Groq, or Mistral key, now keeps segment timestamps, so the Transcript tab and SRT/TXT/JSON/MD export light up automatically.

• In-person meetings get speaker labels — Recordings with everyone in the room and no call used to finish with no speakers at all.

• A redesigned "Coming up" sidebar — Day cards with a today marker, attendee avatars with RSVP state, and a "Join & take notes" button.

• New note directly inside a folder — The same hover "+" that spaces already had.

🎙 Dictation, fixed where it mattered

• If dictation has been failing since 1.8.2, this fixes it. Every dictation configured for OpenAI realtime transcription — the default — failed outright with an Unsupported realtime token provider error. Provider routing now has a single source of truth.

• Local transcription stops thanking you for watching. whisper.cpp's anti-hallucination thresholds are now sent on every request, cutting hallucinated tails from 2.25% to 0.06% across the reporter's 4,814 real dictations.

• Your phone is no longer mistaken for your microphone. A phone offered through Continuity was auto-selected over your real mic and produced empty transcripts.

• A blank model reply never eats your text, and stray tags stop leaking into your titles and notes.

🐧 Linux

• Push-to-talk on Wayland — Hold-to-dictate now works on Hyprland, KDE, and GNOME 48+.

• Hyprland's Lua config is supported — On 0.55+ (and Omarchy), your shortcut no longer vanishes on reload.

• Auto-paste works on non-QWERTY layouts, and dictated text stays on the clipboard when automatic paste can't land.

• Punctuation hotkeys bind on GNOME and KDE, Ptyxis and GNOME Console paste correctly, KDE paste stops stalling up to 19 seconds on a stale portal session, and Sway overlays keep your text field focused.

🪟 Windows

• Dictation pastes into the window you recorded from, even if something else steals focus while transcription runs.

• Whisper models work under non-ASCII and redirected profiles — a CJK or Cyrillic user name used to crash local transcription.

• Packaged builds stop printing logs to your terminal (use --console-logs if you want them).

🛠 Fixes & polish

• GPU acceleration — 1.8.3's multi-GPU regression is fixed: Vulkan pins to your discrete GPU instead of integrated graphics, NVIDIA detection finds nvidia-smi off PATH, and packs cleared during the upgrade are announced instead of silently deleted. A downloaded pack now engages with or without the env flag, and the "GPU available" banner only appears when cleanup actually runs locally.

• AI models — Retired Groq and Tinfoil models are gone from the picker, Gemini 3.5 and 3.1 Flash Lite replace the Flash Lite Google retired, gpt-oss on Tinfoil works again, browsing provider tabs no longer switches your active model, and timeouts and flaky DNS retry instead of failing outright.

• Calendar — Recurring Outlook meetings have names again, personal time blocks stop prompting you to record, Join recognizes Zoom webinars and Teams /meet/ links, all-day events sit on the right day, and a mismatched sign-in fails immediately instead of spinning for two minutes.

• Notes — Cancelling an upload actually stops the work, the chat panel no longer covers what you're reading, generated titles lose their stray quotes, and mirrored Markdown survives line breaks in titles, locked files, and folder names containing ...

• Elsewhere — Turning off App updates now actually stops update checks, a wedged background service can't spin at 100% CPU forever, and zh-Hans / zh-Hant system locales get the Chinese interface.

🙏 Thanks

Thanks to @boseq, @clarkmaan, @hsusul, @IdrisGit, @samkawsarani, @xAlcahest, @xlurie, @Trevor-K-Smith, @Chadpiha, @greatcoat and @stantheman0128 for their work on this release — and a warm welcome to first-time contributors @xpipko, @AdityaPainuli, and @Biotech-Basti! 🎉

Prebuilt Windows mic listener binary for event-driven microphone detection.Uses WASAPI IAudioSessionManager2 to monitor capture sessions with per-PID refcounting across all capture endpoints (hot-plug aware). Announces "CAPABILITY PID" on stdout; PID exclusions are handled by the JavaScript caller.

# OpenWhispr 1.8.3 > Note: 1.8.2 was never published — this release includes everything from it. If you're updating from 1.8.1, all of the below is new to you.

🌟 Highlights

🖊 Edit highlighted text by voice

Select text in any app, trigger the voice agent, and say what you want changed — "make this more formal", "fix the grammar", "turn this into bullet points". The agent rewrites the selection in place instead of pasting a new block. Works even in apps whose accessibility trees stay dormant (Arc, Chrome, Slack, VS Code, Claude Desktop, and more).

🖥 Screen context for the voice agent

Turn on "Share screen context" (Settings → AI Models → Voice Agent, off by default) and the agent sees the display you're working on when you speak — so "reply to this email" or "explain the error on screen" just works. You can route screenshot-carrying commands to a dedicated vision model, and a screenshot that can't be sent never costs you the command. Screenshots live in memory for one request only — never written to disk, stored, or logged.

📅 Microsoft & Apple Calendar

Outlook and Apple Calendar join Google Calendar: your meetings are detected, reminders fire before they start, and meeting notes link to the right event. Google Calendar sync also now handles large calendars completely.

✨ Also new (from 1.8.2)

• Collaboration is free — Any signed-in account can create and join team spaces, share individual notes, and sync shared content. Paid workspaces see prorated seat costs before inviting.

• Managed Enterprise AI — Organizations can centrally configure Amazon Bedrock or Azure OpenAI for cleanup, voice agent, notes, and translation. Employees sign in with company SSO — no cloud keys to handle — and prompts go directly to the organization's cloud account.

• Organization policy, fully enforced — Policy now applies across the whole app: restricted options are hidden with safe fallbacks, and enforcement covers modes, providers, features, sharing, and retention.

• Tinfoil realtime meeting transcription — Tinfoil joins the realtime meeting providers.

• Agent failures are no longer silent — If the agent can't process a command, you get an "Agent Unavailable" notice instead of your raw words pasting into whatever app you were in.

⚡ GPU acceleration you can trust (new in 1.8.3)

• "GPU acceleration active" now means it. The indicator reflects what the transcription engine is actually running on — ready, activating, active, or "could not be activated" with a Retry — instead of turning green whenever a download finished.

• Enable GPU works instantly. Downloading or removing a GPU pack applies immediately — no app restart, no silent CPU fallback.

• A GPU failure never costs you a dictation. If the GPU engine crashes, the same recording is transcribed on CPU and pasted, and the failed backend is remembered instead of being retried on every launch.

• GTX 10-series (Pascal) cards can now use CUDA. The CUDA pack ships Pascal kernels for the first time; cards the build can't run on (Maxwell and older) are offered the Vulkan pack that works — no more loading the model and crashing at first use.

• GPU packs can no longer corrupt each other. Each pack installs into its own directory, installs are atomic (a power cut can't leave a half-installed pack), and old installs are healed automatically.

🔒 Your keys and audio stay where you point them (new in 1.8.3)

• Custom endpoints fail closed. A missing or invalid custom URL for speech-to-text or AI cleanup now fails with a clear error instead of silently sending your audio, prompt, and API key to OpenAI. ⚠️ Action required: if you use the Custom provider and never changed its pre-filled URL, set a real endpoint under Settings — or switch to the OpenAI provider.

• Keys are scoped and stored securely. Your cleanup key no longer rides along to other endpoints, and custom-endpoint keys moved from plaintext into the OS secure store (migrated automatically).

• Every provider remembers its model. Switching providers and back — for transcription and all six AI scopes — restores your previous choice instead of resetting it.

• Transcription errors tell the truth. A broken transcription engine now reports a real error and keeps the recording for retry, instead of blaming your microphone with "No Audio Detected."

🛠 Fixes & polish

• Dictation — No more clipped first words, realtime streaming waits for your final words instead of racing a timer, voice activity detection is now opt-in, dropped transcript segments are retried and surfaced, and transcripts are no longer replaced unexpectedly.

• Meetings & speakers — Speaker labels you set stay put across live and offline identification, live speaker identification lands on Windows, your own dictation no longer triggers meeting detection, diarization writes land on the right note, meeting prompts are steadier, and realtime streams stop cleanly. Intel Mac meetings no longer fail at startup.

• Notes & sync — Cloud sync can't overwrite your local edits with an empty copy, deleted notes can't resurrect, edits made mid-save stay pending, and idle collaboration sync backs off politely.

• Launch at login on Linux — With correct behavior across GNOME and KDE, joining macOS and Windows (which also got start-hidden fixes).

• macOS — The Globe hotkey no longer also opens the system emoji/input switcher.

• Linux — Meeting notifications stay clickable, and the text monitor builds correctly against AT-SPI2.

• Interface — Markdown exports get correct timestamps, empty states close their gaps, the "Coming Soon" badge translates again, and hotkeys with left/right modifiers parse correctly.

🙏 Thanks

Thanks to @Chadpiha, @hsusul, @xAlcahest, @greatcoat, @boseq, and @stantheman0128 for their work on this release — and a warm welcome to first-time contributors @iSparsh and @edwin-luu! 🎉

Prebuilt Linux text monitor binary for auto-learn correction monitoring.This binary uses AT-SPI2 to detect text field changes after paste.

Prebuilt Windows fast-paste binary for clipboard paste and selection-copy operations.Uses Win32 SendInput API with automatic terminal detection for paste (Ctrl+V/Ctrl+Shift+V) and copy (Ctrl+C/Ctrl+Shift+C).

# OpenWhispr 1.8.1 Everything from 1.8.0 — team spaces, web note sharing, new local models — plus a critical fix. If you installed 1.8.0, please update right away.

🚑 Fixed in 1.8.1

• Note content could leak between notes. In 1.8.0, switching from one note to another could copy the second note's content into the first, even if you didn't edit either one. This is fixed — edits now stay with the note they belong to. (If you used 1.8.0, it's worth a quick look through recently-viewed notes.)

✨ What's new (from 1.8.0)

• Team spaces — Create shared spaces alongside your private notes, invite teammates by email with roles, and jump straight in from a one-click link. Your notes stay on your device and sync in the background, with access checked on the server so only the right people can see them.

• Share notes on the web — Publish a note to notes.openwhispr.com for anyone with the link, anyone in your company's email domain, or just the people you invite.

• Workspace settings — Manage members, team spaces, and billing from the new Settings → Workspace section.

• New local AI models — Added the Gemma 4 family of on-device models for faster, fully private AI.

• Privacy retention controls — Choose how long saved recordings and transcripts are kept — down to a single day — and OpenWhispr cleans up the rest automatically.

• See your raw transcript — History now shows the full, unedited transcript with a one-click copy button.

• Dictionary from the command line — List and bulk-update your custom dictionary through the CLI.

🛠 Fixes & polish (from 1.8.0)

• More accurate dictation — Local transcription no longer chops words in half, long recordings don't time out early, and a dropped microphone no longer loses what you just said.

• Steadier cloud & streaming transcription — Sturdier large-file uploads, cleaner NVIDIA Parakeet and AssemblyAI streaming, and your transcript is kept even if the cleanup step comes back empty.

• Safer notes sync — Cloud sync will no longer overwrite a note you've edited locally with an empty copy.

• Better exports — Meeting participants are included, subtitle timing is fixed, and speaker-less segments are merged more cleanly.

• Translation — Local translation is more reliable and handles empty results gracefully.

• Windows — No more console windows flashing during dictation, meetings, and GPU checks.

• Linux — More reliable pasting and meeting detection, and better behavior on Wayland.

• macOS — The Dock icon now follows the control panel, the tray icon toggles it, and the Accessibility setup handles the app not being listed yet.

🙏 Thanks

Thanks to @Chadpiha, @hsusul, @xAlcahest, @IdrisGit, @greatcoat, @dpersek, and @sochotnicky for their work on this release — and a warm welcome to first-time contributors @stantheman0128, @dreasan, @fabrizio2210, @William-Ger, @ambbone, @AIalliAI, and @B3rK-3! 🎉

# OpenWhispr 1.8.0 Work together in shared team spaces and share notes on the web — plus new local AI models, privacy retention controls, and a big reliability pass across dictation and notes.

✨ What's new

• Team spaces — Create shared spaces alongside your private notes, invite teammates by email with roles, and jump straight in from a one-click link. Your notes stay on your device and sync in the background, with access checked on the server so only the right people can see them.

• Share notes on the web — Publish a note to notes.openwhispr.com for anyone with the link, anyone in your company's email domain, or just the people you invite.

• Workspace settings — Manage members, team spaces, and billing from the new Settings → Workspace section.

• New local AI models — Added the Gemma 4 family of on-device models for faster, fully private AI.

• Privacy retention controls — Choose how long saved recordings and transcripts are kept — down to a single day — and OpenWhispr cleans up the rest automatically.

• See your raw transcript — History now shows the full, unedited transcript with a one-click copy button.

• Dictionary from the command line — List and bulk-update your custom dictionary through the CLI.

🛠 Fixes & polish

• More accurate dictation — Local transcription no longer chops words in half, long recordings don't time out early, and a dropped microphone no longer loses what you just said.

• Steadier cloud & streaming transcription — Sturdier large-file uploads, cleaner NVIDIA Parakeet and AssemblyAI streaming, and your transcript is kept even if the cleanup step comes back empty.

• Safer notes sync — Cloud sync will no longer overwrite a note you've edited locally with an empty copy.

• Better exports — Meeting participants are included, subtitle timing is fixed, and speaker-less segments are merged more cleanly.

• Translation — Local translation is more reliable and handles empty results gracefully.

• Windows — No more console windows flashing during dictation, meetings, and GPU checks.

• Linux — More reliable pasting and meeting detection, and better behavior on Wayland.

• macOS — The Dock icon now follows the control panel, the tray icon toggles it, and the Accessibility setup handles the app not being listed yet.

🙏 Thanks

Thanks to @Chadpiha, @hsusul, @xAlcahest, @IdrisGit, @greatcoat, @dpersek, and @sochotnicky for their work on this release — and a warm welcome to first-time contributors @stantheman0128, @dreasan, @fabrizio2210, @William-Ger, @ambbone, @AIalliAI, and @B3rK-3! 🎉

# OpenWhispr 1.7.6 Dictation translation, audio import with speaker detection, GPU support for AMD and Intel, live streaming transcription — and a deep round of self-hosted and reliability fixes.

New

• Translation mode — dictate in any language and paste the text in another. Its own hotkey, up to 5 target languages, and a dedicated model, configured under Settings → AI Models → Translation.

• Import audio from URLs and in batches — paste a YouTube or direct audio link, queue up to 50 URLs and files, and turn on Speaker detection to label who said what. Detection runs on-device by default.

• Vulkan GPU acceleration — the one-click local Whisper GPU flow now covers AMD Radeon and Intel Arc/integrated GPUs on Windows and Linux, with automatic CPU fallback.

• NVIDIA Nemotron streaming models — live text as you speak from a persistent local stream, and dictation now commits the streamed text the moment you stop (no second decode, roughly half the CPU per dictation).

• Liquid AI LFM2/LFM2.5 — five new local reasoning models for on-device cleanup, down to a 0.25 GB model that runs on modest hardware.

• Collapsible sidebar — collapse it for full-width notes; hover the left edge to peek.

• Meeting prompts, unified — calendar reminders now use the in-app overlay (they survive Focus/Do Not Disturb and never appear in screen shares), with one-click Join & transcribe when the event has a meeting link.

Fixed

• Self-hosted: audio uploads, History retries, note formatting, and Chat all respect your self-hosted server now — no more silent fallback to a cloud provider — plus model lists work without /v1 and "disable thinking" works on Ollama.

• Local AI: starting a local cleanup model no longer freezes the whole machine — the llama server runs with a bounded context, and a model that truly doesn't fit fails with a visible error.

• Cleanup providers: Mistral works as a custom provider instead of failing with a 422, Groq no longer fails silently, and if cleanup ever fails your dictation is pasted raw with a toast instead of being lost.

• macOS: pausing your music during dictation works again on macOS 15.4+, and Fn+Arrow shortcuts no longer trigger stray push-to-talk recordings.

• Microphones: a muted or vanished mic falls back to the default device instead of recording silence, and your selection survives Chromium's device-ID rotation.

• Windows: Parakeet installs no longer hang behind an old PATH tar, and the event-driven mic listener that powers meeting detection now actually ships.

• Linux: .deb installs and upgrades no longer fail in containers — and upgrading no longer deletes your downloaded models.

• Notes: search works in every language (Cyrillic, CJK, Arabic, accented text), and Generate Notes regenerates auto-assigned titles while leaving yours alone.

Plus many more — full list in the changelog.

Thanks

Thanks to @xAlcahest, @Chadpiha, and @sgrimbly for their work on this release — and a warm welcome to first-time contributors @hsusul, @sochotnicky, @cgkades, and @Serhii-Leniv!

Prebuilt Windows mic listener binary for event-driven microphone detection.Uses WASAPI IAudioSessionManager2 to monitor capture sessions. Supports --exclude-pid to filter out OpenWhispr's own mic usage.

# OpenWhispr 1.7.5 New models and providers, multiple hotkeys per action, and around a dozen fixes across meetings, notes, transcription, and every platform.

New

• Latest cloud models: OpenAI GPT-5.6 (Sol/Terra/Luna) and Anthropic Claude Fable 5 and Sonnet 5 — plus Claude Fable 5 for AWS Bedrock.

• OpenRouter is now a built-in provider, with a searchable, grouped picker for its 300+ models.

• Corti can run AI cleanup and reasoning, not just transcription. For healthcare use, that text stays on Corti (EU) or the HIPAA-compliant OpenWhispr Cloud — never a third-party model.

• Tinfoil confidential transcription now covers uploaded audio, not just live dictation.

• Multiple hotkeys per action — bind more than one key to dictation, agent, voice agent, or meeting.

• Enterprise / AWS Bedrock: region-aware model IDs, a live model catalog, and Agent Mode (tool-calling) support.

Fixed

• Meetings: remote speakers no longer collapse into one; long calls reconnect past the 60-minute limit without losing audio; fewer words dropped during cross-talk.

• Notes: enhancement no longer reuses another note's transcript or renames notes you've already titled.

• Cloud transcription: long recordings retry failed chunks instead of failing or leaving gaps.

• macOS: dictation into Claude Desktop and claude.ai no longer breaks after the first paste.

• Windows: fixed a false "No audio detected", the firewall prompt for local Parakeet, and local transcription failing on some builds.

• Linux: fixed Wayland global-hotkey crashes on Node 24.

Plus several smaller fixes — full list in the changelog.

Thanks

Thanks to @xAlcahest, @danielmccannsayles, @IdrisGit, and @Alexey-Sachko for their work on this release — and a welcome to @Alexey-Sachko on their first contribution!

# OpenWhispr 1.7.4 A feature-and-reliability release: two new privacy-first ways to transcribe and reason, native Windows meeting audio, sync that keeps working in the background, and a broad sweep of fixes across dictation, clipboard, updates, and Linux — plus security hardening. Nothing here changes your existing setup; it all just works better.

Highlights

• Tinfoil — confidential cloud (BYOK). A new provider that runs both transcription *and* AI cleanup/agent inside attested secure enclaves, so your audio and text stay private even when processed in the cloud. (#875, #944)

• Azure AI Foundry / Azure OpenAI speech-to-text. Point the custom transcription provider at your own Azure deployment — OpenWhispr now speaks Azure's endpoint format and auth. (#997)

• Native Windows meeting audio. Meetings now capture sound from every app on every output device instead of only your default speakers, via a native Windows helper. Older Windows automatically falls back to the previous method. (#960)

• Enterprise SSO at sign-in, right from onboarding. (#1034)

Sync

• Runs in the background, even from the tray — dictations upload and other devices' changes pull down on window focus, on network reconnect, and every 5 minutes, even if you never open the main window. (#1070, #1072)

• Snippets now sync across your devices, like your dictionary already does. (#1037)

• No more duplicate folders when you sign in on a new device. (#1086)

Dictation & notes

• Cleaner cleanup — your transcript gets tidied instead of occasionally getting "answered," and dictations only go to your voice agent when you actually address it by name. (#1073)

• Turkish snippets — triggers containing İ / ı now match correctly. (#1050)

• API keys no longer vanish when you type one and click away — a papercut that mostly hit Corti/BYOK onboarding. (#1039)

Clipboard, audio & media

• Rich clipboard formats survive auto-paste, clipboard restore is faster, and pasting reliably targets the right app on macOS. (#1020, #1038, #1000)

• Mic falls back to your default device when a saved one goes stale, and empty recordings no longer crash transcription. (#978, #891)

• Paused music/video resumes the instant you stop recording (not after transcription) and won't get stuck paused on a quick tap. (#1030, #1061)

Platform & reliability

• Updates: fixed a crash on the first "Install & Restart" click after an update — macOS installs no longer stall. (#1012)

• Local transcription: GPU model reloads after sleep, multi-GPU machines pick the right card, the voice-activity model now ships in packaged builds, realtime streaming is hardened, and llama.cpp was updated. (#1032, #1018, #1000, #1044, #995)

• Hotkeys: modifier-only hotkeys (dictation, voice agent, agent, meeting) now work simultaneously on Windows/Linux. (#1001)

• Linux: installs on openSUSE and launches on hardened kernels that restrict user namespaces. (#1014, #1042)

• Self-hosted: transcription endpoints can now choose their own model. (#1043)

• Quieter startup — suppressed benign Qdrant warnings. (#1083)

Security

• Cleared all critical/high dependency alerts and restored lockfile integrity with a CI guard (SOC 2 secure-code hardening). (#1051, #1069)

Contributors

Thanks to everyone who shipped 1.7.4 — with a warm welcome to our first-time contributors (*): @xAlcahest · @Chadpiha · @IdrisGit · @danielmccannsayles* · @zachdotai* · @bikramjitk* · @SyntaxSawdust* · @UmutEmreOnder* · @greatcoat* · @chirag127*

Prebuilt Windows system audio helper binary for meeting transcription.Captures system audio via WASAPI process loopback (exclude mode), hearing every application on every output device while excluding OpenWhispr's own audio. Requires Windows 10 2004 or later; older systems automatically fall back to Chromium display-media loopback.

✨ New

• Snippets — say a shortcut like "cal link" and the full text gets pasted

• Voice Agent hotkey — one key sends your speech straight to your AI assistant

• Medical dictation via Corti for clinical-grade accuracy

• Smarter setup that adapts to how *you'll* actually use OpenWhispr

• Redesigned dictionary that now syncs across your devices

• Audio upload gets its own settings and a cancel button

• Discarded dictations are saved, so nothing gets lost

• Notification controls to pick which alerts OpenWhispr can show

• New AI models: Claude Opus 4.8, Gemini 3.5 Flash, Gemma 4, plus xAI transcription

# OpenWhispr 1.7.2 A small patch on top of 1.7.1 — zero unnecessary Keychain prompts on first launch, cloud transcription working again on Electron's net.fetch, and the Note Formatting selector now actually controls model routing.

Features

• Note Formatting selector now routes Generate Notes. The Note Formatting tab's model / provider / mode selectors were a no-op end-to-end in 1.7.1 — every Generate Notes call silently fell through to the Cleanup model. Now wired through the same per-scope plumbing, with a Cleanup fallback so untouched-settings users keep their current behavior.

• Wave Terminal paste. Wave Terminal is now in the Linux terminal allowlist, so auto-paste routes through Ctrl+Shift+V instead of the default Ctrl+V.

Bug Fixes

• No more spurious macOS Keychain prompts on first launch. First-launch Keychain prompts drop from ~3 to 0. The secret-crypto backend no longer eagerly probes itself before any window appears — it defers Keychain access until you actually save your first secret. The safeStorage key backup is also only written when a new master key is generated, not on every launch.

• Cloud transcription works again. The 1.7.0 migration from https.request to Electron's net.fetch carried over a manual Content-Length header, which Electron rejects as a forbidden Fetch header — failing with net::ERR_INVALID_ARGUMENT before any bytes hit the wire. Fix covers all five upload paths: cloud transcribe, chunked cloud transcribe, retry, file upload, and BYOK whisper-compatible.

• Notes view state stability. Fixed a stale-ref issue where switching between notes could lose unsaved enhanced-content edits.

Features

• macOS mouse-button hotkeys for dictation

• Google Calendar: new "sync primary calendar only" toggle to skip shared ones

• Make automatic note titling optional in Settings

• Smarter local transcription that ignores silence and breathing, fully configurable for local models

Bug Fixes

• Cloud meeting transcription restored after a recent OpenAI Realtime API change

• Local Whisper transcription restored on Windows

• Music auto-pause and resume works again on the latest macOS

• Self-hosted models now use their own URLs and API keys instead of falling back to cleanup settings

• Custom dictionary now reaches meeting note cleanup

• Language preference preserved on retry, in meetings, and in the dictation preview

• Custom prompts sync across windows instantly without restart

• Transcription sync no longer crashes on bad cloud data

• Idle local AI models release GPU memory after a timeout

• Works correctly behind corporate proxies for model downloads, GPU detection, and calendar sync

• Linux: Pop!OS COSMIC support, GNOME Wayland terminal auto-paste, Konsole paste fix, and the dictation overlay no longer steals focus on Sway/i3/Hyprland

Breaking for Mac users: please delete your old OpenWhispr first. 1.7.0 uses a new app ID — we had to migrate Apple Developer accounts, so you'll be asked to re-grant permissions on first launch. Notes, settings, API keys, and downloaded models carry over automatically.

Summary

• More sign-in options — new: Sign in with Microsoft, Sign in with Apple on macOS. Auth runs on our own infra now (migrated to Better Auth); self-hosters can point at their own server.

• Sessions in your OS keychain — survive crashes and Electron restarts. No more random sign-outs.

• API keys encrypted at rest — all 12 BYOK + enterprise creds moved from .env to the OS keychain. Silent one-time migration on first launch.

• Background meeting recording — navigate to other notes, open Settings, switch tabs; the audio pipeline keeps going. New floating pill shows live mic levels and clicks back to the recording note.

• Per-note diarization preferences persist across stop/resume.

• Cleaner mic capture with a new acoustic gate + better echo cancellation. Music pause/resume on Windows works again.

• Per-scope LLM setup — pick different providers/models for cleanup, agent, formatting, and chat. You can now have text cleanup toggled ON but your dictation agent toggled OFF.

• NVIDIA Parakeet parakeet-unified-en-0.6b — new English-only model, 5.91% avg WER, ~631 MB.

• openwhispr CLI talks directly to the desktop app via a local HTTP bridge (127.0.0.1, bearer-token); falls back to cloud when the app is closed.

• Proxy-aware fetches + OS CA trust — corporate TLS interception works.

• Auto-learn corrections now work for Cyrillic, CJK, Arabic, Devanagari.

• A bunch of other small bug fixes and performance improvements.

Detailed changelog

• build(mac): migrate Apple signing to Gizmo Labs Inc. by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/639

• fix(qdrant): set cwd to STORAGE_DIR so packaged app doesn't crash on startup by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/640

• feat(sync): cross-device delete propagation by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/646

• feat(meeting): interop cloud streaming providers + AEC/VAD improvements by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/656

• Re-grant permissions modal for 1.6.11 upgraders by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/659

• fix(chat): persist first user message when creating a new conversation by @xAlcahest in https://github.com/OpenWhispr/openwhispr/pull/662

• fix(tls): trust OS CA store for Node-side TLS; keep realtime WS on OpenAI only by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/657

• Fix adding non-ASCII languages to dictionary by @wake0up0ne0 in https://github.com/OpenWhispr/openwhispr/pull/666

• Fix dictionary correction tracking for TextPattern-based controls (alongside existing ValuePattern support) by @wake0up0ne0 in https://github.com/OpenWhispr/openwhispr/pull/665

• Bundle ID migration: com.herotools.openwispr → com.gizmolabs.openwhispr by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/669

• feat(cli): local HTTP bridge for unified CLI by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/676

• fix(notes): update folder selection when moving notes between folders by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/678

• refactor: split dictation cleanup from dictation agent + per-scope LLM config by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/677

• refactor(auth): switch desktop to Better Auth + add Microsoft sign-in by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/686

• fix(network): proxy-aware fetches + actionable connectivity errors by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/687

• chore(release): 1.7.0 confidence cleanup by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/689

• ci(release): rename VITE_NEON_AUTH_URL → VITE_AUTH_URL by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/690

• fix(auth): initiate desktop OAuth from browser, not renderer by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/692

• feat(auth): Sign in with Apple on macOS by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/691

• fix(onnx): cap embedding segments and isolate inference in utility process by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/693

• fix(sidecars): reap children on quit and on next launch (#683) by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/694

• docs(changelog): plain-English 1.7.0 + lockfile refresh by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/695

• fix(llama): raise Vulkan startup timeout, stop server before re-download by @xAlcahest in https://github.com/OpenWhispr/openwhispr/pull/698

• fix(media): fall back to media key when GSMTC fails on Windows by @xAlcahest in https://github.com/OpenWhispr/openwhispr/pull/697

• docs(readme): acknowledge Hugging Face as model hub by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/703

• feat(transcribe): generate clientTranscriptionId for cloud sync dedup by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/702

• feat(security): encrypt API keys at rest via safeStorage by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/629

• fix(media): use AsTask bridge for WinRT async calls in GSMTC scripts by @xAlcahest in https://github.com/OpenWhispr/openwhispr/pull/706

• fix(media): resume playback on no-audio-detected event by @kdenney in https://github.com/OpenWhispr/openwhispr/pull/701

• feat(auth): switch desktop to Authorization: Bearer + safeStorage by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/704

• fix(ci): compile macos-media-remote binary in release and build workflows by @kdenney in https://github.com/OpenWhispr/openwhispr/pull/700

• feat(reasoning): self-hosted OpenAI-compatible parity + thinking-mode toggle by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/708

• feat(parakeet): add parakeet-unified-en-0.6b model by @milanleonard in https://github.com/OpenWhispr/openwhispr/pull/713

• feat(meeting): background recording with floating pill and responsive layout by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/709

• fix(runtime): unblock model downloads, surface ONNX worker errors, dedupe hotkey re-register by @gabrielste1n in https://github.com/OpenWhispr/openwhispr/pull/716

• fix(linux): resolve symlink in launcher wrapper by @xAlcahest in https://github.com/OpenWhispr/openwhispr/pull/717

New Contributors

• @wake0up0ne0 made their first contribution in https://github.com/OpenWhispr/openwhispr/pull/666

• @kdenney made their first contribution in https://github.com/OpenWhispr/openwhispr/pull/701

• @milanleonard made their first contribution in https://github.com/OpenWhispr/openwhispr/pull/713

Features

• Speaker diarization controls — per-meeting toggle, "N others in call" stepper, auto-label on 1-on-1s, fallback to "You/Others" when labeling is off

• Integrations hub — new top-level view with MCP connectivity for Claude, ChatGPT, and Cursor; API keys moved here

• Enterprise connections - connect to your organisations instance via bedrock

Bug Fixes & Improvements

• Settings reorganization — "AI Models" split into Speech-to-Text and Language Models with clearer sub-tabs

• Agent hotkey moved under Hotkeys

• Meeting transcription tuned to reduce GPT-4o hallucinations during silence

• Model cache opens at the correct folder so downloaded models are actually visible

• Sync — cleared transcripts and deleted folders no longer reappear after restart

• Echo leak detector now runs pre-AEC so it can actually see system audio bleed

New features

Transcript Export — Save transcripts to disk as TXT, SRT, or JSON directly from the meeting view. Cloud Sync — Notes, folders, conversations, and transcriptions now sync bidirectionally across devices. API Keys — Create and manage API keys from Settings to integrate OpenWhispr with your workflows programmatically. Linux Push-to-Talk — Native push-to-talk via evdev with guided permission setup. Auto-Paste Toggle — New setting to disable automatic pasting after dictation. Agent Folders — The agent can now list, create, and match folders semantically when organizing notes.

Improvements

• Upgraded to Electron 41 and Node 24

• Redesigned provider tabs as compact pill buttons

• Local model descriptions replaced with clickable spec links

• llama.cpp detection now probes /v1/models and prefers /chat/completions

Fixes

• Windows hotkey no longer false-triggers after Win+L lock/unlock

• Fixed stuck recording when push-to-talk key-up is missed on Windows

• Relaxed speech gate thresholds with no-audio toast for failed transcriptions

• Retry button now shown on failed transcriptions

• Notes sync when updated externally

• Tuned speaker diarization to prevent excessive speaker creation

• TLS/certificate errors now surfaced in model download UI

• Floating icon position persists across restarts

• Fixed notification window sizing and dismiss circle visibility

New features

• Meeting recording with on-device speaker diarization — capture mic and system audio together, and every segment is automatically labeled with who said it. Diarization runs 100% locally: voices never leave your device. System audio capture now works on Windows and Linux (Chromium loopback and the native audio portal), not just macOS.

• Local live transcription preview — see your words appear in real time as you speak, streamed from a local Whisper or Parakeet model before you release the hotkey. Fully on-device, no cloud round trip.

• Self-hosted support — first-class option to point OpenWhispr at your own LAN server for both transcription and agent inference. One click to swap between OpenAI, Anthropic, Gemini, a local GGUF, or your self-hosted endpoint.

• Speaker to contact linking — voice profiles bind to calendar attendees, so recurring participants get named automatically after the first tag.

• Improved echo cancellation — WebRTC AEC now runs as a native sidecar, paired with a custom cross-correlation detector and text-level dedupe to cleanly handle external speakers, system-audio capture, and low-echo-quality mics.

• Gemma 4 models — Gemma 4 31B and Gemma 4 26B MoE added to the local model registry.

• Smart transcription error handling — when a transcription fails, inline CTAs let you retry, switch provider, or open logs.

Linux

• Hyprland global hotkeys via hyprctl, plus correct active-window detection for terminal paste.

• GNOME shortcut conflict detection, active-key display, and proper fallback when the requested shortcut is taken.

Fixes

• Windows: installer DPI scaling corrected, GPU compositing re-enabled.

• Local Whisper: stricter speech gate cuts hallucinated output on silent audio.

• Diarization: session-scoped results, disk spooling for long meetings, reliable spinner, correct speaker label propagation.

• Notes: transcripts scope to their originating note, view auto-switches when a meeting recording starts.

• UI: fixed 420px window width that grows only with text, consistent across phases; preview overlay redesigned.

• Settings: inference mode auto-switches when you select a new provider.

• i18n: speaker labels, cloud settings, integrations, and calendar strings tightened across locales.

• Stability: 22 React Compiler lint issues resolved, tighter effect dependencies, cleaner hook usage across 15 files.

Prebuilt WebRTC Audio Processing sidecar for meeting mic echo cancellation.Binaries built from native/meeting-aec-helper/ against pinned webrtc-apm + abseil-cpp commits.

AI Chat & Semantic Search — Ask questions about your notes with the new embedded chat panel. Conversations sync to the cloud, and a local semantic search engine finds notes by meaning, not just keywords. Meeting Improvements — Calendar attendees automatically appear on meeting notes, auto-detection works for browser meetings, and echo cancellation cleans up mic input. Save Notes as Files — Export your notes to local Markdown files that mirror your folder structure. Responsive Settings — Settings dialog adapts gracefully to smaller windows. Bug Fixes — Resolved issues with meeting transcription, folder switching, clipboard pasting, and participant saving. Improved Linux Wayland support and Windows build signing. Huge thanks to @xAlcahest @DamianPala and others for their contributions to platform stability and improvements!

Rich Text Notes

• Notes now use a rich text editor with Obsidian-style live preview — Markdown syntax hides as you type, giving you a clean writing experience

Meeting Transcription Upgrades

• Dual-channel transcription — mic and system audio are captured separately with speaker-labeled chat bubbles so you can see who said what

• Timestamped segments — meeting transcripts now include timestamps in chronological order

• Smarter meeting notes — AI-generated summaries are now speaker-aware for more accurate meeting recaps

New Local Models

• Mistral Nemo 12B and Gemma 3 12B are now available for on-device AI processing

macOS Audio Improvements

• Native system audio capture via CoreAudio Tap — no more "screen recording" permission prompt on macOS 14.2+

• macOS 15+ now shows the correct system audio consent dialog instead of the legacy screen recording one

Linux

• KDE Wayland support — native global shortcuts now work on KDE Plasma via D-Bus, joining GNOME and Hyprland support

• Fixed KDE Plasma overlay window and hotkey behavior

• Fixed clipboard paste reliability on KDE Wayland

Simplified Permissions

• Permission prompts consolidated to a single "Grant Access" button

• Permissions are now re-checked against the OS each time you open settings, so the UI always reflects your actual state

Bug Fixes

• Fixed Gemini agent streaming routing to the wrong endpoint

• Fixed Windows mic volume being permanently altered during dictation

• Fixed mono transcription failure on Linux

• Fixed Bluetooth audio issues during meetings

• Fixed meeting detection notifications firing when you're already in a meeting

• Fixed held modifier keys not releasing before paste on Windows

• Fixed paused media being unpaused during dictation

• Fixed Google OAuth users skipping onboarding

• Fixed AI cleanup prompt refusing to transcribe command-like speech

• Fixed agent hotkey conflicts not showing a warning