■ New • Meetings: Automatic meeting detection is now available. Whisper Transcription notices when a call starts in apps like Zoom, Microsoft Teams and Slack and offers to record it, so you no longer have to start meeting recordings by hand. • Added the GPT-6.1 Terra AI model. ■ Improvements • App audio recording: If the app you are recording quits during the recording, Whisper Transcription now tells you. A recording that includes your microphone keeps going; one without a microphone stops and keeps what was captured. • Qwen3-ASR: Better language handling for Czech, Danish, Dutch, Filipino, Finnish, Greek, Hindi, Hungarian, Indonesian, Italian, Macedonian, Malay, Persian, Polish, Romanian, Swedish, Thai and Turkish. When you choose one of them, Qwen3-ASR now transcribes in that language. ■ Bugfixes • Fixed: On macOS 26 and later, app audio and meeting recordings could capture system audio at half pitch with some USB headsets that have a built-in microphone, or capture no system audio at all with multichannel audio interfaces such as those from Audient and Universal Audio. • Fixed: App audio recording could capture no audio from the selected app if that app had been quit and reopened before you pressed Record.

■ Improvements • M1 Macs: Qwen3-ASR is now an opt-in, experimental model on M1 Macs, because the Neural Engine on these Macs can stop it partway through longer files. To use it, turn on "Show Qwen3-ASR on this Mac" in Advanced Settings, under Transcription Settings, Parakeet / WhisperKit / Qwen3-ASR. If you already downloaded it, it stays available and is marked as experimental. Turning it off switches you to another downloaded model and offers to delete its files. ■ Bugfixes • Fixed: The progress percentage for WhisperKit transcriptions could jump back and forth with voice activity detection (VAD) turned on.

■ Improvements • M1 Qwen mitigation: When the Neural Engine stops a transcription partway through, MacWhisper now tries again once automatically. • M1 Qwen mitigation: We're still working on Qwen3-ASR stability on M1 Macs. This release adds more detail to exported logs, so if you run into problems, please send us your logs by choosing Export Logs from the Help menu. ■ Bugfixes • Fixed: A Neural Engine error was reported as "Model cache corruption", which deleted the model cache and made the next transcription recompile the model, taking a few minutes. • Fixed: A file opened just as MacWhisper was launching could occasionally fail to transcribe. • Fixed: Starting or ending a FaceTime or other call during a recording could break the recording's microphone track, leaving a file that couldn't be opened.

■ New • The app is now localised in Spanish, French, Japanese, Simplified Chinese, Swedish, and Finnish. Please let us know if you see any translation errors! ■ Bugfixes • Qwen3-ASR transcription progress now updates correctly. • Old temporary audio files to free disk space.

■ New • Added the Claude Haiku 5.5 AI model. ■ Improvements • Damaged recordings, such as some Microsoft Teams exports with a corrupted stretch, are now read past the damage in one pass instead of after several failed conversion attempts, so they start transcribing sooner. • Whisper C++: Transcribes directly from your original file, without writing a temporary WAV copy first. • Clearer errors: a file MacWhisper can't read now says "Could not open the audio file", a video without sound says "No audio track found", and an empty recording says "No audio to be read", instead of "Operation Stopped" or an error from the engine. • Settings: Removed "Downmix audio to mono before transcription" and "Use Argmax Audio Pipeline" from Advanced settings. All local engines now share one audio decoder that always mixes multichannel audio evenly to mono, so neither option is needed. ■ Bugfixes • Qwen3-ASR now loads on the Neural Engine on M6 Macs. An already downloaded model is patched in place. • Fixed: Local transcription could stop before the end of some MP3 and AAC recordings, with no error. This affected variable-bitrate MP3s without a length header and raw .aac files from voice recorders. • Fixed: WhisperKit could transcribe only part of a file, or nothing, with no error. Examples: only the first second of some screen recordings, Opus audio inside MP4 files, and files whose extension doesn't match their contents. • Fixed: Opus recordings saved with an .mp3 extension (Plaud recorders do this) gave an empty transcript with WhisperKit and were sent to cloud providers unconverted. They are now recognised by their contents and converted like any other Ogg file. • Fixed: Files whose first audio track is disabled, such as a muted commentary or alternate-language track, were transcribed as silence. MacWhisper now transcribes the track you hear when you play the file. • Fixed: Ogg files with an upper-case .OGG extension failed to import on macOS 15.0 to 15.3. • Fixed: A damaged Ogg file showed "This recording is too long to import on this version of macOS" and suggested an update that wouldn't help. It now says the file can't be opened. • Fixed: Temporary audio copies made for transcription could linger in the app's storage instead of being removed afterwards.

■ New • Qwen3-ASR is now available on M1 Macs, and is recommended in the Local Models list on supported Macs. • Low disk space: MacWhisper warns you when disk space runs low during a voice memo or app audio recording, and stops the recording safely before the disk fills up. The warning floats over other apps when the recording window is not in front. • AI models: Added GPT-6.1 Sol. • Ignored segments: Paste a list to add many segments at once, one per line. Reset to Defaults now asks for confirmation. ■ Improvements • System audio recording: On macOS 26 and later, system audio is now captured through CoreAudio instead of ScreenCaptureKit, so it only needs the System Audio Recording permission instead of Screen Recording. • Echo cancellation: Now on for everyone, made possible by the move to CoreAudio above. ■ Bugfixes • Spotlight: Spotlight's daily reindex no longer launches MacWhisper in the background. • Privacy: The Send Anonymous Analytics setting now also covers the speech SDK's telemetry. • App audio: Pressing Start twice while the permission is still settling no longer starts two recordings. • Onboarding: Fixed the welcome sheet growing in from the top-left corner on first appearance.

■ New • Speaker recognition: You can now choose which speaker recognition model you want to use for local file transcription, Nemotron or Pyannote. You can adjust this from speaker settings. Nemotron has the highest accuracy but is limited to 8 speakers. • AI models: Added Sonnet 5.5, GPT-6 Sol, GPT-6 Luna, Grok 4.6 and Grok 4.7. ■ Improvements • ElevenLabs: You can now transcribe recordings up to 10 hours and files up to 3 GB, including with speaker detection enabled, and we’ve added Sinhala language support. • Speaker recognition: Transcript information now shows which speaker recognition model was used. • Watched folders: Wait for files to finish copying before they start processing, download iCloud-only files automatically, and process files inside newly added subfolders when subfolder watching is enabled. • AI services: New services default to Sonnet 5.5, GPT-6 Luna or Grok 4.7, and xAI branding has been updated to SpaceXAI. • Find and Replace: Improved the layout of the replacement list’s import, export and remove-all controls. • File access: Clearer guidance to restart the app when an update or reinstall prevents it from saving access to files and folders. ■ Bugfixes • Fixed: Your microphone speech in live meetings is attributed to your meeting speaker instead of being split into multiple speakers. • Fixed: Watched folders could miss files during replacement or re-transcribe existing files after a folder read error. • Fixed: Watched-folder transcriptions could produce duplicate exports when `Auto-Export` was also enabled. • Fixed: Queue groups could appear out of order, and failed or cancelled batch imports could prevent subsequent files from processing. • Fixed: Crashes related to drag-and-drop imports and floating window layout. • Fixed: Model downloads could fail because of an invalid or revoked Hugging Face token stored on the Mac. • Fixed: Transcripts could be marked as having speaker recognition even when no speakers were identified.

■ New • Speaker detection is now powered by Nvidia's Nemotron 3 Diarization model (Sortformer v3) and recognizes up to 8 speakers per recording. • When more than one transcript is open, the Open section of the sidebar has a Close All button. • AI Services settings have a Test Connection button that checks the service is set up and replying. ■ Improvements • Added Claude Opus 5.5, Gemini 3.7 Flash and Gemini 3.8 Flash to AI Services. New Gemini services default to 3.8 Flash. • Model Memory Usage in Advanced Settings now has separate load and unload settings for File Transcription and Live Captions. • On macOS 27, AI prompts can be reordered by dragging in the Prompts settings. • Transcriptions made from a YouTube or media URL now show their source link in the inspector, with copy and open buttons. • The summary language is now remembered. Its new default, "Same as Transcript", follows the language the transcript was made in rather than the one the engine detected. • The alert shown when the database can't be opened now has a Show in Finder button. ■ Bugfixes • Fixed a crash on quit when a Whisper C++ model was still loaded. • Fixed a crash when transcribing extremely long audio (over about 25 hours) with WhisperKit, Parakeet or Qwen3-ASR. It now shows an error instead. • Restored the Transcribe Audio to Text Shortcuts action for macOS 15 and 26. If your shortcut shows "This action is not supported on Mac", replace the action with the restored one. • Fixed double-clicking a word while a transcription was still in progress seeking to the wrong position. • Fixed Qwen3-ASR speaker detection assigning the first or last word of a sentence to the wrong speaker. • Fixed Qwen3-ASR failing when you manually pick a language it can't be locked to. In that case it now detects the language automatically. • Fixed a downloaded Apple Speech model sometimes disappearing from the model list and being deselected. • Fixed a long URL running underneath the download progress label in the Transcribe URL field. • When a downloaded URL transcribes but the engine fails, the alert now shows the underlying error. • Fixed New Subfolder not showing the new folder's name field the first time. • Fixed the folder sidebar briefly flashing empty when creating a folder.

■ Improvements • Improved checks for the Mac's model requirements before download. • Local Models in Settings hides empty categories when your Mac can't run those models. • Cleaner Local Models list, with a new Qwen provider icon. ■ Bugfixes • Fixed a crash on macOS 15 when opening the Live Captions language menu. • Unfortunately removed support for Qwen3-ASR on M1-series Macs after user feedback.

■ New • Qwen3-ASR: Added local transcription in 30 languages, including support for live meetings and Live Captions. Requires Pro and an Apple Silicon Mac running macOS 26 or later. • Whisper C++: Added voice activity detection to skip silence and a setting to control whether previous text influences the next passage. ■ Improvements • Whisper C++: Updated the transcription engine and simplified advanced settings, including a clearer option to retry low-confidence passages. • Audio import: Improved compatibility with different WAV file formats. • Recording: Reduced the processing overhead of the recording screen and waveform. • Sidebar: Refined spacing and alignment. ■ Bugfixes • Fixed: MacWhisper actions not appearing in Shortcuts and Siri in release builds. • Fixed: Dragging files from some apps could leave imports stuck, including in batch and podcast workflows. • Fixed: File drops being incorrectly accepted by the transcript editor. • Fixed: Obsidian exports failing when note titles contained slashes or colons. • Fixed: Hidden transcription models appearing in model search results. • Fixed: Incorrect sizing of the Roadmap view in Settings.

■ Improvements • Recording: smoother waveform rendering and reduced processing overhead. ■ Bugfixes • Fixed: dragging recordings from apps such as Voice Memos could get stuck, including in Batch and Podcast mode. • Fixed: dropping a file onto the transcript editor could replace the displayed transcript with the file’s contents. • Fixed: Obsidian exports could fail when note titles contained slashes, colons, or other unsupported characters. • Fixed: a layout-related crash on macOS 27. • Fixed: the Roadmap layout and Manage Export Formats button styling in Settings.

■ New • Sharing: Send audio directly to Whisper Transcription from other apps using the macOS Share menu, including Voice Memos. • AI models: Added GPT-5.6 Sol, Terra and Luna, GPT-6 Astra, and Claude Fable 5.1. ■ Improvements • Folders: Expand and collapse individual folders, with your layout remembered between launches. • Folders: Use the arrow keys to expand and collapse folders, and hover over folders while dragging to open them. • Translations: Improved translations throughout the app. ■ Bugfixes • Fixed: Failed retranscriptions could replace the original transcription. Originals are now preserved until retranscription completes successfully. • Fixed: Permanently deleting voice memos and dictations could leave associated audio files behind. • Fixed: Deleted transcriptions could still appear in dictation history or affect sidebar counts. • Fixed: Renaming folders from the context menu could fail to start editing. • Fixed: Opening shared files and reopening the main window could behave unreliably.

■ New • Apple Intelligence support lets you ask Siri AI questions about your transcripts as well as search for their content with natural language • Shortcuts: On macOS 27, transcribe audio or video with your choice of model, language, speaker identification, and whether to save the result to your library. • Spotlight and Siri: Find and open transcriptions through Spotlight and Siri on macOS 27, with searchable transcript content and support for referring to the transcription currently on screen. • Shortcuts export: On macOS 27, retrieve transcript text and pass transcriptions between actions for saving, sharing, previewing, and formatted export. • Folders: Organize your library into folders, move items with drag and drop or context menus, and see item counts in the sidebar. • Drag and drop: Drag a transcription into Finder to export a .whisper archive, or into a text destination to insert its transcript. • Live meetings (Pro): Follow a meeting transcript as you record, with live speaker identification. Enable Live Meeting Transcription in Advanced Settings → Recording Settings and select a compatible model, such as Parakeet or WhisperKit. • Live meeting controls (Pro): Rename and merge speakers, highlight the active speaker, keep the transcript window on top, and open or end the meeting from the menu bar. • AI integrations (Pro): Send AI summaries to connected integrations from the summary tab, and choose the content sent by your automations. • Improved dictionary import • Media timecodes: Automatically detect embedded start timecodes in supported MOV and MP4 files, show the start timestamp in file information, and apply the offset in the editor. ■ Improvements • Accessibility: Improved VoiceOver labels and navigation throughout onboarding, history, transcript details, dictation setup, and settings. • History search: Find transcriptions by their original filenames and search for exact phrases in quotation marks, with matching preview highlights. • File imports: Import a mixture of media, transcripts, and models together, with clearer reporting when individual files cannot be imported. • Transcription queue: See errors for individual tracks, open available transcripts from partially completed imports, and clear completed groups while keeping unfinished work. • Settings search: More responsive filtering with less repeated work while building the settings search index. • Notion summaries (Pro): Preserve headings and lists when sending summaries to Notion, with improved handling of long Chinese, Japanese, and Korean text. • Integration feedback (Pro): Clearer success and failure reporting when sending content, with consistent send menus for transcripts and summaries. • Library protection: Automatically create rotating database backups before migrations • Sidebar: Cleaner folder and section layouts, hidden zero counts, and context-menu actions to move items into or remove them from folders. ■ Bugfixes • Fixed: Transcription cancellation and model unloading could interfere with later jobs or leave the queue stuck. • Fixed: Transcription could remain stuck during speaker identification without surfacing the available transcript. • Fixed: WhisperKit model loading could be unreliable. • Fixed: Missing source audio now produces an early error instead of being passed to the transcription engine. • Fixed: Some database changes could be lost when the task saving them was cancelled. • Fixed: Copying or exporting transcripts without speaker identification could lose paragraph breaks. • Fixed: History search could mishandle quotation marks or display outdated results; failed searches now offer Retry. • Fixed: Live Captions could fail to load Pro models despite an active license. • Fixed: Retranscription recovery could leave replacement sessions hidden. • Fixed: Editing an AI prompt title did not always save the change. • Fixed: ElevenLabs server errors could incorrectly appear as network errors.

■ Improvements • Models you download in Settings are now prepared right after downloading, so your first transcription with a new model is no longer held up by a one-off preparation step • Transcription failure alerts and notifications now name the engine that failed • Model lists are fetched far less often, which cuts network traffic at launch and speeds up cold starts on a poor connection ■ Bugfixes • MacWhisper now tells you when macOS refuses access to the login keychain • Deleting an AI service now also removes its stored API key from the keychain • Fixed a crash when an audio conversion was cancelled while it was still reading the file • Custom AI services no longer fail when the provider returns a non-streaming response, and rejected requests now show a clearer error • Groq services that used a retired chat model now migrate to a current one automatically, instead of failing every request with "Selected AI Service Not Available" • A flaky network could produce a wrong model list with no error shown; MacWhisper now falls back to the last known-good list • Cancelling a URL download no longer tears down the progress display of a newer download started in the same field • The floating playback bar no longer covers the end of the transcript in the editor

■ New • Cloud transcription: Added Cohere Transcribe and Speechmatics, including model, language, and region options. • Deepgram privacy: Added an option to opt out of Deepgram’s Model Improvement Program; it is enabled by default. ■ Improvements • AI models: Updated Google Gemini and xAI Grok models, with automatic migration from retired model selections. • Export options: Added speaker-name, timestamp, and millisecond controls to batch exports, watched-folder presets, custom presets, and the export sidebar. • Cloud transcription: Improved reliability, language handling, and error messages across providers. • Provider setup: Added dedicated icons for Cohere, Speechmatics, Corti, Cartesia, Avalon, and Volcengine. • Queue: Failed items now show clearer errors without misleading progress. ■ Bugfixes • Fixed: YouTube downloads no longer fail with HTTP 403 or produce empty files. • Fixed: Watched folders now ignore unsupported file types. • Fixed: Transcript exports no longer show unsupported timestamps, drop lines, or add stray spacing in word-grouped output. • Fixed: Speaker names in word-grouped exports now appear once per speaking turn. • Fixed: Empty transcript segments no longer appear in OpenAI, Deepgram, and Gladia results. • Fixed: ElevenLabs now uses the correct language code for Tagalog. • Fixed: Exported HTML transcripts now use the correct text color.

■ Improvements • Gemini models now show friendly names instead of raw API identifiers • Model lists no longer re-download from Hugging Face on every network change • Refreshed the Pro tab in Settings ■ Bugfixes • Downloaded WhisperKit models' first load now works without a network connection

■ New • Transcript editor: create and assign a new speaker directly from the “Assign Selection to” menu • ElevenLabs: choose US, EU, India or Singapore data-residency endpoints for Enterprise accounts ■ Improvements • Assistant transcription: broader language support • URL transcription: more reliable YouTube and direct-media downloads, smarter stream selection and clearer error messages • Search: AI-generated titles and short summaries are now searchable • Transcript editor: small transcripts load faster and switching between transcripts does less background work • History: smoother scrolling and faster day grouping • Model loading: improved local-model stability • Licensing: clearer guidance when entering a free-version license ■ Bugfixes • Fixed: home search could display “No Results” even when matches were found • Fixed: history screens could occasionally render an empty list • Fixed: switching transcripts could send content, search or playback updates to the wrong editor • Fixed: “No text available” was misaligned in history rows

■ New • Added OpenAI gpt-transcribe as a cloud transcription option, with live streaming output and regional Chinese support • Redesigned model picker showing your 3 most recently used models • New settings sidebar icons and reorganized settings categories • Custom Obsidian properties support • Claude Opus 5 model support for AI features ■ Improvements • Opening transcriptions is much faster • Added missing German and Dutch translations • Diarization models are now grouped under a Speaker Identification section in the model pickers ■ Bugfixes • Fixed Apple Speech models that would disappear after relaunch • Better support for DeepL rate limits when translating long transcripts • Fixed export for free users in an edge case • Fixes a potential crash on macOS 26.0 when starting a transcription

■ Improvements • Fixed an issue where some pro users could not use the summary feature • When exporting the transcript format you can now also choose to export it to markdown • Improved the design of the info page in settings and added a toggle for enabling automatic updates • Clearer errors when models can not be downloaded

■ Improvements • Live transcription now streams in order and groups into readable paragraphs as you speak • The transcript inspector now opens by default and remembers whether you left it open • Speaker and timestamp columns stay hidden until speaker identification finishes, so the transcript doesn't jump around • Ogg files now convert natively where macOS supports it, and very long files report an error instead of crashing • Refreshed the available Anthropic models for AI features ■ Bugfixes • Fixed retranscribing a recording with a different language or model appearing to do nothing - the new transcript, language, and model now appear immediately • Fixed the home screen sometimes showing "No transcription available" for a recording that had just finished transcribing • Fixed recordings showing content from a deleted transcript - they now correctly show as untranscribed, and search no longer matches deleted transcripts • Fixed relative date headings like "Today" in history becoming wrong after midnight • Fixed Ogg files not being importable on macOS 14 • Fixed text export repeating the speaker's name on every line • Fixed the Converting/Decoding progress icon being clipped • Fixed Select All (Cmd+A) not working on AZERTY and other non-QWERTY keyboard layouts

■ New • Added a microphone picker to the menu bar, making it easy to quickly switch your preferred microphone or open microphone settings ■ Improvements • Transcriptions now start faster by loading the model while preparing the audio • Cancelling a transcription while audio is being prepared is now much more responsive • Long sections from the same speaker are automatically divided into more readable paragraphs • Copying a selection containing multiple speakers now includes their names • Improved Settings opening performance and search highlighting • Exported diagnostic logs now include macOS crash reports to help us investigate issues • Improved transcription loading states and progress labels • Clarified that the Voice Activity Detection setting only applies to WhisperKit models • Improved the position of the delete button in the transcription history ■ Bugfixes • Fixed missing spaces when pasting copied transcripts into Microsoft Word • Fixed an issue where imported .whisper files could have unplayable audio when the original media extension was missing • Fixed media exports sometimes using the wrong file extension • Fixed microphone settings not updating when changed elsewhere in the app or system

■ New • New Editor View! You can now (finally) edit your transcript directly in the main editor view. Dramatically improved performance, easily add new paragraphs, quickly re-assign speakers, favorite any part of the transcript you want and a lot more! The old transcript view is still shown for translations, for now. Let us know what you think and if there's anything we can improve by emailing us at support@macwhisper.com! • Added Gladia as a new Cloud transcription provider. You can use it with your own API key. ■ Improvements • Redesigned AI services screen in settings to be clearer and a lot nicer looking • Updated the newest AI models for most popular AI providers • Localization improvements • Show the copy button in full screen translation mode • Add filler words toggle for Deepgram • Added the option to choose the region for Deepgram. You can choose between US and EU (thanks Daniel) • Expanded Deepgram supported languages for Nova 3 • Improved playback when you have multiple transcriptions open • When using Cloud transcription providers, the files are no longer loaded into memory when sending them to the service ■ Bugfixes • Prevent a Documents permission prompt on first launch • The menubar icon now remembers it's position if you changed it (thanks Raphael) • Added a potential fix for an issue where the app could use a lot of CPU while idle. If you run into this still please let us know. • Improved sync between audio and video scrub bars • Fixed an issue where macOS would return an error about an active FaceTime call when trying to record a meeting in an edge case (thanks Andree) This version is the last one to support macOS 14. Future updates will require macOS 15 or newer. You can still keep using this version for as long as you want. You can also reach out to support@macwhisper.com for help.

■ Improvements • Made it easier to enable the new Editor display mode from the top right of the transcript view. Please give it a try and let us know what can be improved on support@macwhisper.com • Improved the colors of the Editor display mode in dark mode • Added more padding in the Editor mode

■ New • Editor display mode (Beta) - a brand new transcript display mode that lets you freely edit your transcription like a text document. It's in beta and we'd love your feedback! To try it out for now you can turn it on from Advanced settings. • System-wide keyboard shortcut to start a new voice recording ■ Improvements • Redesigned the voice memo recording screen • Added a back button to return Home from an open transcript • Select previously used speakers for podcasts to quickly reassign hosts you've used before • Edit a transcription's title directly from the sidebar • Added a "no verbatim" option for ElevenLabs transcription that cleans filler words out of the transcript • Audio-only MP4 files are now supported • Clearer error messages when a YouTube download fails • All in-progress translations now appear in the sidebar • When a model's cache is corrupted, MacWhisper now suggests switching models or turning off Fast Loading • Word-level timestamps are now preserved in exports when a start-time offset is set ■ Bugfixes • Fixed a bug where the recording timer display could pause • Fixed speaker name colors not appearing in HTML and PDF exports • Removed Japanese from Parakeet's supported languages (it wasn't actually supported) • Fixed the back button after a voice memo transcription incorrectly returning to the recording screen • Fixed a layout glitch when toggling the sidebar

■ Improvements • Huge performance improvements when scrolling the Transcript screen while playing back audio • Clearer warning when your microphone is unavailable because the laptop lid is closed (clamshell mode) • Other performance improvements ■ Bugfixes • Fixed a crash when recording with microphones that use interleaved audio (e.g. Blue Yeti) • Fixed recordings getting stuck and impossible to stop after closing the laptop lid. With the default ScreenCaptureKit backend, the recording now ends cleanly when the lid is closed and can still be transcribed. With the CoreAudio backend, recording resumes when you reopen the laptop (no audio can be recorded when the lid is closed) • Handle multiple taps on the Stop Recording button correctly • Restored the Live Captions button in the Home screen shortcuts • Fixed duplicate formats appearing in Batch Export