Fixed in 0.8.9

A small follow-up to 0.8.8 for clearer AI setup and saved-key handling.

• Turning AI off keeps saved API keys. Remove an unused key explicitly in AI Setup.
• Use Apple Intelligence for Dictation & cleanup with Default AI set to None. Summaries, chat, and Transforms require another provider because Apple's model has a small context window.
• See each task's provider or Off. Prompt settings follow the provider assigned to that task.

Previously selected Apple Intelligence? Check Settings → AI Setup after updating. An old Apple default is cleared; an old Apple analysis override falls back to Default AI. Saved keys remain. Choose Apple Intelligence under Dictation & cleanup to use it there.

Also included from 0.8.8

• Retry failed dictations from History, with more reliable Fn shortcuts and rapid restarts.
• Edit transcripts and saved AI results, regenerate in place, and browse a faster Library.
• Choose separate AI models for cleanup and analysis, and choose from more output languages.
• Try the optional Orukeet speech-model preview and Nemotron automatic speaker detection. Additional model downloads may be needed.

Requires Apple Silicon and macOS 14.2+. Apple Intelligence requires macOS 26+ and enabled system models.

0.8.9 release notes and CLI compatibility · Full 0.8.8 notes

MacParakeet 0.8.8

Retry failed dictations, edit transcripts and saved AI results, and browse your Library faster.

Dictation

• Retry failed recognition from History when Save dictation history is enabled.
• More reliable Fn shortcuts and rapid restarts, with fixes for cancelled takes interfering with new recordings. Stop no longer waits on stalled live recognition.
• Optional capture sounds, top or bottom pill placement, and a choice to let Escape pass through. Clean-mode dictation understands spoken “question mark” and “exclamation mark.”
• An optional AI-polish shortcut for individual dictations. Requires AI Formatter to be enabled with a configured provider.

Library and editing

• Edit transcript passages and saved AI results. Switching result tabs preserves your draft, and saving protects against overwriting newer edits.
• Regenerate in the existing result tab; cancelling restores the saved result.
• Faster Library loading and scrolling, more responsive long transcripts, and more room for Notes. Manage Prompts from the Library toolbar.
• Cancellable folder imports and safer recording recovery.

AI and speech models

• Choose separate AI models for cleanup and analysis, and choose from more languages for AI results. Run saved Transforms from the menu bar.
• Use on-device Apple Intelligence on eligible Macs with macOS 26+. Best suited to short rewrites and cleanup.
• Automatic speaker detection now uses Nemotron, with up to eight speakers per source. An additional model download and setup may be needed before offline use. Explicit speaker counts continue to use Community-1.
• Try the optional Orukeet speech-recognition preview with a separate model download. Parakeet v3 remains the default; Orukeet does not support live dictation preview or recognition-time custom vocabulary. @Nathan-Roll1 contributed the MacParakeet integration for Oruk’s model.
• Four-step onboarding with dictation practice and bundled CLI 4.9.0.

Requires Apple Silicon and macOS 14.2 or later.

Full release notes, model requirements, and compatibility details

Sorry — 0.8.6 broke hold-to-talk (human error). But not for long — here's 0.8.7. Hold-to-talk is restored.

• Hold Fn (or your PTT shortcut) when the mic is already granted. The first-time system sheet still waits for the next hold. (#1096, #1097)
• Fn works with Caps Lock on. (#1099)
• Hands-free cancel/stop and the Undo pill sit tighter. Hold-to-talk keeps the wider pill. (#1098)
• First-run can continue without a microphone. Dictation still asks before you speak. (#1070)
• AI Formatter stays off on new installs, with separate prompts for transcripts and dictation. If you already turned it on for transcripts, that stays on. (#1093)
• Optional streaming cursor types finished dictation at the caret. Off by default in Settings → Dictation. Paste remains the normal path. (#1092)
• Connect Moonshot (Kimi), DeepSeek, Qwen, Z.AI, or MiniMax as first-class AI providers. (#1087)
• On macOS 14 (Sonoma), Parakeet’s encoder stays off the Neural Engine so long files are less likely to crash. (#1089)
• macparakeet-cli is 4.4.0: the same AI provider settings as the app, plus history rename/favorite and speaker corrections.
• Wrapping up after stop was only a status label. The recording was already saved. The tile now moves on. (#1079)
• Start meetings muted, Clean English “um” (turn it off in Vocabulary for Portuguese or German), and optional preserved discarded dictations are still here.

Skip the microphone if you only transcribe files. Keep AI Formatter off until you want it. Optionally type finished dictation at the caret.

• First-run can continue without a microphone. Dictation still asks before you speak. (#1070)
• AI Formatter stays off on new installs, with separate prompts for transcripts and dictation. If you already turned it on for transcripts, that stays on. (#1093)
• Optional streaming cursor types finished dictation at the caret. Off by default in Settings → Dictation. Paste remains the normal path. (#1092)
• Connect Moonshot (Kimi), DeepSeek, Qwen, Z.AI, or MiniMax as first-class AI providers. (#1087)
• On macOS 14 (Sonoma), Parakeet’s encoder stays off the Neural Engine so long files are less likely to crash. (#1089)
• macparakeet-cli is 4.4.0: the same AI provider settings as the app, plus history rename/favorite and speaker corrections.
• Wrapping up after stop was only a status label. The recording was already saved. The tile now moves on. (#1079)
• Start meetings muted, Clean English “um” (turn it off in Vocabulary for Portuguese or German), and optional preserved discarded dictations are still here.

The Transcribe tile could stay on Wrapping up after you stopped a meeting. That was a status label only — the recording was already saved. 0.8.5 clears it.

• Wrapping up was cosmetic. Capture and transcription were already finished. The tile now moves on to Transcribing, then idle. (#1079)
• Turn on Start meetings muted to begin recording with your microphone off. Unmute from the live meeting panel when you want to speak. System-audio-only capture ignores it. Off by default.
• Clean dictation and file transcripts now strip standalone English “um” along with uh, umm, and uhh. Meeting transcripts stay verbatim. If you dictate in Portuguese or German, turn off Also remove “um” in Vocabulary — it is on by default, and um is a real word in those languages.
• Preserve discarded dictations keeps cancelled dictations in History, marked Cancelled, instead of deleting them. Nothing is pasted, including Paste Last. Off by default, and it needs Save dictation history.
• If live transcription is off, the Transcript seed-of-life no longer keeps spinning. Leaving Settings while a speech engine is still loading no longer leaves a stuck spinner.
• Library recordings without artwork share a Seed of Life cover: seven circles on a night field. Each meeting still looks a little different. YouTube and video thumbnails stay as they are.
• Skip auto-record for one calendar meeting, or a repeating series, without turning auto-start off. Record, the hotkey, and the menu bar still work.
• Turn off live transcription while you record if you want the Mac to work less during the call. It stays on by default, and the full transcript still runs after you stop.
• A stuck or slow microphone no longer blocks the rest of the recording. If you quit during that startup, you can End & Transcribe or Discard.

Thanks @gtfierro for the live transcription setting.

Meetings can start muted, Clean drops English “um,” and cancelled dictations can stay in History if you want them.

• Turn on Start meetings muted to begin recording with your microphone off. Unmute from the live meeting panel when you want to speak. System-audio-only capture ignores it. Off by default.
• Clean dictation and file transcripts now strip standalone English “um” along with uh, umm, and uhh. Meeting transcripts stay verbatim. If you dictate in Portuguese or German, turn off Also remove “um” in Vocabulary — it is on by default, and um is a real word in those languages.
• Preserve discarded dictations keeps cancelled dictations in History, marked Cancelled, instead of deleting them. Nothing is pasted, including Paste Last. Off by default, and it needs Save dictation history.
• If live transcription is off, the Transcript seed-of-life no longer keeps spinning. Leaving Settings while a speech engine is still loading no longer leaves a stuck spinner.
• Library recordings without artwork share a Seed of Life cover: seven circles on a night field. Each meeting still looks a little different. YouTube and video thumbnails stay as they are.
• Skip auto-record for one calendar meeting, or a repeating series, without turning auto-start off. Record, the hotkey, and the menu bar still work.
• Turn off live transcription while you record if you want the Mac to work less during the call. It stays on by default, and the full transcript still runs after you stop.
• A stuck or slow microphone no longer blocks the rest of the recording. If you quit during that startup, you can End & Transcribe or Discard.
• A one-word speaker mix-up, or a short unlabeled gap, follows the surrounding speaker when both sides match. Fewer tiny extra bubbles; same speakers.

Thanks @gtfierro for the live transcription setting.

Library recordings without artwork now share a Seed of Life cover: seven circles on a night field. Each meeting still looks a little different. YouTube and video thumbnails stay as they are.

• Skip auto-record for one calendar meeting, or a repeating series, without turning auto-start off. Record, the hotkey, and the menu bar still work.
• Turn off live transcription while you record if you want the Mac to work less during the call. It stays on by default, and the full transcript still runs after you stop.
• A stuck or slow microphone no longer blocks the rest of the recording. If you quit during that startup, you can End & Transcribe or Discard.
• A one-word speaker mix-up, or a short unlabeled gap, follows the surrounding speaker when both sides match. Fewer tiny extra bubbles; same speakers.

Thanks @gtfierro for the live transcription setting.

What's New in 0.8.2

• Skip auto-record for one calendar meeting. Skip this meeting, or a repeating series, without turning auto-start off for everything else. You can still record it yourself from Record, the hotkey, or the menu bar.
• Live transcription during recording. You can turn this setting off to use less processing during the meeting. It stays on by default, and the full transcript is still created after you stop.
• A slow microphone no longer blocks the rest of the recording. MacParakeet can still record everyone else on the call even if your microphone is stuck or slow to start. If you quit during that startup, you can End & Transcribe or Discard.
• Fewer tiny speaker bubbles. A one-word speaker mix-up, or a short unlabeled gap, now follows the surrounding speaker when both sides match. This cleans up extra tiny bubbles without changing the speakers MacParakeet detected.

Thanks

Thanks to @gtfierro for the live transcription setting.

What's New in 0.8.1

• Bring existing recordings into Meetings. Import an audio or video file without changing the original, or split a saved meeting into separately transcribed parts.
• Correct timed transcripts. Edit or merge transcript lines, then undo, redo, or reset your corrections. The corrected text stays consistent in playback, search, AI context, exports, meeting files, and the CLI.
• Set up Microsoft 365 calendars. Connect Microsoft 365 and Exchange calendars through macOS Calendar for meeting detection, without a separate cloud sign-in inside MacParakeet.
• Keep Bluetooth capture running through silence. MacParakeet no longer treats valid silence from an established microphone as a stalled input. System Default also gets a fresh startup attempt before falling back from a Bluetooth microphone.
• More dependable transcription. Long Parakeet jobs on macOS 14 avoid unsafe concurrent Neural Engine work, speaker-count limits remain in effect throughout diarization, and YouTube anti-bot errors now include practical next steps.
• CLI 4.1.0. The bundled CLI adds commands for meeting import, splitting, and timed transcript corrections.

What's New in 0.8.0

• More dependable recording and meetings - improved recovery across microphone, Bluetooth, AirPods, and audio-route changes; safer finalization and back-to-back meetings; and saved meeting notes that autosave and can inform post-meeting prompts. (#857, #860, #869)
• Richer transcripts - edit speaker names and corrections, use rich Markdown in results, chat, and Live Ask, select embedded audio tracks, and export W3C DAPT transcripts. (#839, #854, #957, #960)
• A more useful Library - create, rename, recolor, and apply labels with stable colors; choose a persistent grid or list layout; see clearer favorite and audio state; and give recordings without artwork a generated cover. (#958, #987, #988)
• Better prompts and AI controls - manage Transcript Prompts and Live Ask separately, organize prompts in collections, use reliable after-meeting automation, and set model-aware generation controls per prompt. (#961, #968, #990, #991, #992)
• More control - choose whether Discover can fetch its feed, control menu-bar visibility, and manage Custom Words in bulk. (#876, #937, #978)

Thanks

Thanks to @alfred-sa, @robd003, @keyanfayaz, @adambbutler, @tsahli, @jeffpoulton, and @stormj-UH for their contributions. Read the full release notes.

What's New in 0.7.3

• More reliable dictation starts - MacParakeet follows the macOS System Default microphone again, and prepares the speech model while idle to reduce cold-start delay and clipped opening words.
• Clearer speech engine control - choose one primary speech engine, with an optional Advanced override for final meeting, file, media, and URL transcription.
• Safer meetings - fixed a verified stop-time crash, strengthened start/stop/recovery, restored the Meetings panel, and clarified optional transcript auto-save. (#792, #807)
• Better transcription and imports - Whisper shows progress while its model prepares, Nemotron timed transcripts preserve word spacing, local imports keep their original filenames, and slower yt-dlp metadata startup is allowed. (#805, #806, #809)
• Other fixes - current Claude models no longer receive a deprecated temperature parameter, and Clean processing preserves multilingual um. (#793, #774)

Thanks

Thanks to @dudemeister23 for separate dictation/transcription engine routing and dictation cold-start improvements. (#783, #785)

• More local speech options - Cohere Transcribe is now available as an optional local batch speech engine, with safer memory checks and a GPU/ANE compute setting.
• Voice Return phrases - add multiple trigger phrases, including multilingual equivalents, while still only triggering Return from phrases spoken at the end of dictation. (#607)
• Cleaner meeting transcripts - MacParakeet can reduce speaker bleed by using a cleaned microphone track for final meeting transcription, while preserving the original recording as fallback.
• More recoverable meetings - finalization status is visible, failed processing can be retried, source audio cleanup is safer, and meeting artifacts now clearly identify mic, system, mixed, and cleaned audio.
• Better speaker and meeting controls - speaker detection is easier to manage, speaker renames stay in sync, a new Settings option can hide the meeting recording pill, and Bluetooth-output setups can prefer the built-in microphone for dictation input when appropriate. (#648)
• Library and transcript upgrades - bulk transcript export, local transcription title renaming, in-transcript find, improved reading controls, and cleaner PDF/DOCX exports. (#657, #647)

Fixes & Reliability

• Improved dictation tail handling, including hold-to-talk final-word cases; the Parakeet v3 short-utterance/final-phrase issue remains open. Fixed a macOS 14 Neural Engine crash from concurrent speech inference. (#562)
• Added CLI retranscription for saved Library records. (#627)
• Improved onboarding permission recovery, meeting Markdown refreshes, legacy meeting artifact recovery, CLI health diagnostics, and release verification for bundled meeting echo assets.
• Redacted user paths and filenames from diagnostic logs.

Thanks

Contributions this release from @dudemeister23 for Cohere Transcribe, the Cohere compute setting, and related recovery polish; and @saphid for release/version safeguards and keyboard test fixes. Thank you both.

• Parakeet Unified (English) — a new selectable Parakeet speech model with native streaming dictation, punctuation, and capitalization. Choose it under Settings → Speech Model. English-only, with no word-level timestamps, so the default v3 model is still best when you need timed transcripts or other languages. (#520)
• Meeting source modes — record from microphone and system audio together (the default), microphone only, or system audio only. Microphone-only needs no Screen Recording permission. (#540)
• Meeting audio retention — keep a meeting's source audio forever (the default), auto-delete after a set number of days, or remove it right after transcription. Your transcript, notes, and AI results are always kept.
• Back-to-back meeting recording — start the next meeting without waiting for the previous one to finish transcribing. (#535)
• Configurable dictation undo window — keep the 5-second cancel countdown, shorten it to 1 second, or turn it off. (#404)
• Automatic meeting titles — when an AI provider is configured, new meetings get a concise generated title instead of a timestamp. (#546)
• A sacred-geometry finish for saved meetings — when a recording is saved, the pill blooms through a Metatron's Cube before settling into a checkmark. (#587)

Improvements & fixes

• Multi-select Library cleanup — select several items and delete them at once, or remove just a meeting's stored audio while keeping the transcript, notes, and exports. (#498)
• Vocabulary corrections now apply to meeting transcripts.
• Dictation history notes expand inline for longer dictations. (#543)
• The timed transcript view now disables editing instead of switching to raw text. (#570)
• Plus a clearer "microphone unavailable" message, a fix for invisible live transcript text, durable meeting artifact folders after audio deletion, corrected meeting permission copy across source modes, and stability and privacy hardening.

Thanks

Contributions this release from @mvanhorn (configurable dictation undo window #583, and a PATH-probe security fix #589) and @hungryhenry101 (invisible live transcript fix #574). Thank you both.

• Live dictation preview — your words now roll in above the dictation pill as you speak. It's display-only — the pasted text still comes from the final, accurate transcription. This has been turned on by default, so if you prefer the original/minimal pill, you can toggle it off anytime under Settings → Capture → Dictation. This is only supported for Parakeet and Nemotron (no Whisper).
• New English speech engine (Beta) — Nemotron Speech Streaming EN 0.6B, a lighter (~600 MB) English-only option with live streaming dictation.
• Meeting auto-stop (Beta) — optionally stop a recording after prolonged silence or when the meeting app closes, with a countdown you can cancel. Off by default.
• Meeting audio retention — choose whether to keep or auto-delete a meeting's source audio.
• Stop & delete in-progress meeting transcriptions.
• Streamlined first-run — a faster, dictation-focused onboarding that prepares your speech model sooner.
• Cleaner meeting transcripts, steadier mic handling across device changes and Bluetooth, clearer export errors, and assorted fixes.

Thanks

Contributions this release from @athurdekoos (Nemotron English engine and fixes) and @andreagrandi (AI Formatter options; he's also built an Obsidian plugin that syncs meeting notes via the CLI). Thank you both.

• Paste (almost) any link — X, Vimeo, TikTok, Instagram, Facebook, Twitch, SoundCloud, Apple Podcasts, and any other site yt-dlp supports. The card recognizes the platform and shows its mark as you paste. (#466)
• Podcasts — Apple Podcasts links resolve straight to the episode audio via iTunes; from the CLI, transcribe --podcast "Lex Fridman episode 400" does a freetext search. (#463)
• New engine: Nemotron 3.5 (Beta) — an opt-in fast multilingual on-device engine, alongside Parakeet and Whisper. Pick it under Settings → Speech Recognition (one-time model download). (#464)
• Instant dictation (opt-in) — keeps the mic warm so dictation starts instantly and doesn't clip your first words. Off by default; macOS shows the mic indicator while it's on. (#414)

Improved

• Timed view for raw transcripts — see word-level timing even without clean processing. (#458)
• Transform shortcuts moved to ⌃⌥1 / ⌃⌥2 / ⌃⌥3 — Control-Option avoids the Option+digit special-character conflict for Polish, Distill, and Decide. (#453)
• Feedback diagnostics — the optional bug-report capture log is now more focused and still contains only device state and counters — no audio, no transcript text. (#451)

2026-06-06T09:14:02.118Z  dictation_diagnostics_session_start  version=0.6.22 build=20260609180743
2026-06-06T09:14:48.512Z  dictation_capture_start              permission_status=3
2026-06-06T09:14:48.560Z  dictation_capture_first_buffer       sr=16000.0 ch=1 frames=4800 has_float_data=true
2026-06-06T09:14:54.004Z  dictation_capture_stop               duration_s=4.875 max_rms=0.071234 non_silent_buffers=95
2026-06-06T09:14:54.330Z  dictation_transcribe_complete        chars=49 words=9 engine=parakeet
        

• Multi-select dictation cleanup — Turn on Select Multiple in your Dictations history to select and delete several entries at once. (#425)
• Transcript context for meeting AI — AI summaries and chat now include transcript timestamps and speaker labels, not just flat text. Toggle Rich (default) or Plain in AI settings. (#393, #431)
• Inline insertion style for Clean dictation — The new Inline option drops the leading capital and trailing period, so dictated text slots into the middle of a sentence. Sentence stays the default. (#413)
• Opt-in diagnostic log for bug reports — When you send feedback about a microphone or capture issue, you can optionally attach the capture-diagnostics log. It contains device state and counts only — no audio or transcript text — and the form previews exactly what's in it.

Fixed

• Live meeting Ask no longer carries transcript context over from a previous meeting or a stale AI-context setting.
• More reliable Transforms loading under rapid edits.

• Hide the unfinished app-aware AI Formatter profile controls. This is coming soon.
• Simplify the AI Formatter dictation settings.

• Added app-aware AI Formatter smart defaults for common writing contexts like messaging, email, browsers, notes, documents, code, and terminal.
• Added CLI support for transcribing media URLs supported by yt-dlp, plus speaker-count and speaker-range controls for diarization.

Refined

• Made AI formatting for live dictation opt-in while keeping AI formatting available for file and meeting transcripts.
• Added a confirmation step before switching speech engines, since first-time Whisper setup can take several minutes.
• Refined the Capture settings workflow selector and improved Meetings workspace responsiveness.

Fixed

• Fixed a Meetings Intelligence badge layout issue where text could wrap into vertical letters.

• Reverted the v0.6.17 microphone routing change after it caused silent dictation or meeting recordings on some Macs; we’ll revisit the fallback behavior in a follow-up release.
• Improved resize and navigation responsiveness in the Meetings workspace.
• Kept Parakeet as the default onboarding speech engine when English is among your preferred macOS languages.

• Address cases where Bluetooth headphones could switch into low-quality call audio during meeting recording when using the macOS System Default microphone.

Model Storage Cleanup

• Added controls to delete downloaded speech recognition models you are not using, so you can reclaim disk space and download them again later. #311

New

• Batch file transcription - Drag and drop multiple local audio/video files, or a folder, and MacParakeet will process them as one batch. Batches run sequentially, continue if one file fails, and support up to 200 supported files at a time. Batch queueing in the Mac app applies to local files and folders; YouTube transcription remains one URL at a time. Short demo in #397.
• Completion sound for transcriptions - The new "Notify when transcription finishes" setting plays a sound when file, YouTube, or batch transcription completes. When MacParakeet is in the background, it can also use macOS notifications for a completion banner. #397
• English-only Parakeet model option - Settings now lets you choose between multilingual Parakeet TDT 0.6B v3 and English-only Parakeet TDT 0.6B v2. v3 stays the default; v2 is useful for English-first users who want to avoid language auto-detection mistakes. #311, #398

New

• Added a new Meetings workspace that brings calendar controls, recording setup, live Ask, and configurable post-meeting prompt cards into one place.

Refined

• Improved meeting live previews with voice activity detection, so transcript updates better follow natural pauses during conversation.
• Reduced CPU usage while meeting recording is active, especially for the floating recording pill.
• Clarified the first-load caption so users know the initial speech engine setup can take a bit.

Fixed

• Fixed dictation sometimes clipping the beginning of speech when "Pause media while dictating" was enabled. #383
• Fixed a bug where starting dictation from the on-screen pill and stopping with Fn could lose the audio/transcription. #379

• MacParakeet now supports a dark mode toggle in Settings. The app looks especially nice in dark mode :) #376

• Fixed Calendar access in the signed macOS app, so Calendar setup now works from onboarding and Settings. #368
• Improved AI model pickers and feedback reports: model pickers better reflect the models available from your configured provider, and feedback reports can now include up to five screenshots. #365
• Added an opt-in setting to keep dictated text on the clipboard after paste, useful for remote desktop and slower paste targets. Thanks @rhynelabs #362

• Added automatic media pause/resume for dictation, so supported media like YouTube, music, or podcasts pause while you speak and resume when capture stops. #351
• Improved dictation hotkeys so push-to-talk and hands-free can share the same custom trigger, like Right Command. #350
• Added a meeting recording mic mute control that keeps system audio recording while muting only the local microphone track. #309
• Fixed microphone permission recovery after reinstall/TCC reset, refreshed the macOS app icon, and updated the README screenshots. #354

Thank you Matt for the generous sponsor - a lot of serotonin was released for many hours and there is still an afterglow :) And thank you to @rhynelabs for the PR contributions!

New

• Calendar reminders and opt-in auto-start are now available, with a short countdown before recording begins. Recordings still stop manually.
• Added native macOS menus for capture, navigation, window controls, updates, and help.

Refinements and fixes

• Improved dictation shortcuts with single-tap hands-free mode, side-specific modifier shortcuts, duplicate shortcut blocking, and clearer Fn guidance.
• Text snippets can now be edited in place, and Vocabulary, Feedback, and Library surfaces are easier to scan.
• Meeting and transcription views are clearer, including meeting names in rows, a more visible New Transcription action, and more accurate engine status copy.
• LM Studio setup now supports optional API tokens and handles local endpoint limits more gracefully.

• Transform hotkeys now paste the result into the currently focused input, so text selected from read-only places like browsers, PDFs, and terminal scrollback can still be transformed and pasted where you are writing.
• Added Paste Last Transform and Recent Transforms to the menu bar for quickly reusing recent Transform results.

• Fixed a 0.6.7 regression where dictation could be saved as pasted without appearing in Terminal, terminal apps, or some web chat inputs. #303

• Transforms — Rewrite selected text anywhere on your Mac with configurable hotkeys and your chosen LLM provider.
• Transform History — Review, search, copy, and delete saved Transform runs.
• Playback speed controls — Adjust retained audio and video playback speed.

Improvements

• Improved retained YouTube audio playback, readable filenames, thumbnail handling, and embedded metadata.
• Improved AI setup discovery, unsaved settings feedback, speech-engine switching feedback, and transcript prompt discovery.
• Improved meeting recording diagnostics and default microphone capture behavior.

Fixes

• Fixed clipboard restoration after dictation.
• Fixed assorted bugs and reliability issues.

• Dictation Stats — Dictations now has a dedicated Stats tab with daily streaks, a heatmap, and top apps so you can see where and how consistently you dictate.
• Meeting audio access — Meeting transcripts now expose the original recording with Show in Finder and Save Audio As, with safer export handling to avoid accidental data loss.

Polish

• Improved Stats tab details, including richer hover popovers, steadier heatmap interactions, and cleaner narrow layouts.

• Fixed a crash loop affecting macOS Sonoma 14.x users after updating to 0.6.4. #268
• Fixed shortcut recording so the key you are trying to assign is no longer swallowed or triggered while recording a new hotkey. #263

• Pause/resume meeting recordings — pause from the floating pill, tile, or panel; the rosette freezes (amber tint), and the file stitches back together on resume. #235
• Modifier-only hotkey chords — bind shortcuts like ⌥⇧ with no letter key required. #234
• Configurable YouTube audio quality — Settings now offers M4A (Apple-friendly default) or "Best available" (YouTube's highest-quality Opus stream). #237
• Split dictation shortcuts — push-to-talk and hands-free now have independent bindings. #185
• Smoother first dictation — proactive speech-engine pre-warm plus an unobtrusive caption while the model loads.

Bug Fixes

• Short hold-to-talk taps no longer drop with "More audio please". #216
• Dictation no longer pastes leftover clipboard text on slower apps. #236
• Settings → Dictation: hotkey labels render cleanly (e.g. ⇧⌘N, 🌐 Fn, 1) instead of duplicated glyph + name.

Under the hood

• Transcription Library scales smoothly with large histories.
• Tighter yt-dlp/ffmpeg lifecycle: child processes always exit, failed downloads always clean up.

• Meeting recordings now keep microphone audio going if macOS interrupts system audio capture. #232
• Transcribing local files and URLs now preserves more media metadata, including artwork and thumbnails, and URL downloads are more resilient. #229
• Speaker detection is still available, but now starts off by default. Turn it on in Settings → Speaker detection to add speaker labels. #230

Sorry about the YouTube transcription failure some fresh installs hit in recent builds. This update fixes the signed helper issue that could break YouTube downloads and repairs affected helper copies automatically.

• YouTube downloads work again — Fixed the bundled yt-dlp helper signing so YouTube transcription can start normally. #226

• Audio Input test reliability — Fixed a 0.6.0 regression where some system-default USB or camera microphones could fail the input test on macOS. #218
• Live transcript readability — Fixed speaker text colors that could be hard to read on light backgrounds. #219

MacParakeet now has three main ways to work: dictate anywhere, transcribe files and URLs, and record meetings locally on your Mac.

• Meeting recording — Capture microphone and system audio, then turn the meeting into a local transcript without sending audio off your Mac.
• Live Notes, Transcript, and Ask — Take notes while recording, watch the transcript come in, and ask questions about the meeting as it develops.
• Clearer permission guidance — If Screen Recording or Microphone access is missing, MacParakeet now points you to the right setup path before recording.
• Optional WhisperKit multilingual transcription — Use a local WhisperKit engine for languages beyond Parakeet's English-first model.
• Cleaner transcript defaults — Speaker diarization is now off by default for simpler transcripts. You can turn speaker labels on whenever you want them.
• More polished setup and transcription — Onboarding, AI provider setup, file transcription, meeting recovery, and update handling all received reliability and usability improvements.

MacParakeet remains local-first: core dictation, transcription, meeting audio, and transcripts stay on your Mac.

Fixes a 0.5.6 regression that blocked users from saving Ollama or LM Studio settings when the server was on a LAN IP, mDNS hostname, Tailscale address, or 0.0.0.0 bind.

• Local providers accept any host again — Ollama and LM Studio no longer require HTTPS or loopback. Any reachable endpoint works.
• Provider-aware placeholder — The base URL field now shows a sensible default for whichever provider is selected.
• Clearer error copy — Validation message no longer mentions HTTPS/localhost for local providers.

Thanks to the user who reported #118.

Two weeks of improvements in one release — new global shortcuts, three more ways to plug in AI, and a stack of fixes including a long-standing first-dictation crash for folks with virtual audio tools installed.

• Transcription hotkeys — Assign global shortcuts for File Transcription and YouTube Transcription in Settings → Transcription. Trigger either flow from anywhere on macOS.
• LM Studio support — Run models locally via LM Studio and pick them straight from the AI Provider menu.
• OpenAI-Compatible provider — Plug into any OpenAI-API-shaped endpoint (vLLM, LocalAI, LiteLLM, llama.cpp server, hosted third parties). API key is optional. Loopback endpoints automatically get local-friendly timeouts.
• AI formatter — New option under AI Provider settings that runs your dictations and file transcripts through your chosen AI model to clean up grammar, punctuation, and paragraphing. Toggle on/off, customize the prompt, reset to default.
• Auto-copy fallback — If an app blocks paste simulation, your dictation now lands on the clipboard automatically so nothing is lost.
• First-dictation crash fixed — Squashed a long-standing crash for users with Krisp, BlackHole, Loopback, or other virtual audio devices installed.
• Quieter when idle — Fixed perpetual SwiftUI animations that kept CPU and GPU awake when you weren't dictating.
• Unstuck Transcribe screen — The Transcribe tab no longer freezes on a completed transcript when you start a new one.
• Hotkey quality-of-life — New "Reset to Default" button in the hotkey recorder, and the option to disable the dictation hotkey if you prefer to trigger from the menu bar only.
• "More audio please" — Redesigned the no-speech state with a softer, friendlier visual (drifting warm leaf + serif italic label) and a warmer message when MacParakeet doesn't hear you speak. Also fixed a menu bar icon that could linger red after dictation ended.
• Smoother launch and shutdown, plus a fix for the Feedback panel's community link pointing at the archived repo.

Huge thanks to @krunkosaurus for contributing LM Studio support, the OpenAI-Compatible provider, and several other improvements this cycle!

Highlights: #93 · #107 · #108 · PR #100 · PR #109

MacParakeet can now distinguish left and right modifier keys when you record bare modifier hotkeys.

• Left/right modifier support — Record hotkeys like Right Option, Left Shift, or Right Command and MacParakeet will respect the exact side you chose.
• Better hotkey reliability — Fixed an edge case where the opposite-side modifier could slip between taps and still count toward a side-specific double-tap.
• Backward compatible — Existing generic modifier hotkeys still work the same way and continue to trigger from either side.

PR #63 · #62