Release notes

Version 6.0 puts names on your transcripts. A new People directory shows who spoke in which meeting, and a preview speaker model can tell remote participants apart. You can also run summaries on an AI server you already have, let other apps react when a meeting is done, and choose where your recordings and Markdown files are saved. Settings has been rebuilt to make room.

Highlights

People: know who said what

People is a new section in the sidebar. Assign a speaker in a transcript to a person, and their name replaces the label everywhere that meeting appears, including summaries and chat. Each person's page lists the meetings they spoke in, and global search (⌘K or Ctrl+K) finds them too.

• Names follow voices. Re-transcribe a meeting and assignments carry over when the match is unambiguous. If it isn't, MeetingDebrief keeps the assignment as it was rather than guess.
• No accidental duplicates. Renaming someone to a name that already exists offers a merge instead.
• Stale summaries are flagged. Naming a speaker marks that meeting's summary as out of date, so you can regenerate it with the right names.
• A clear reason when speakers can't be separated. If a transcript can't tell speakers apart, it now says why, such as a silent microphone, a single audio track or a missing model.

Tell remote speakers apart (preview)

Until now, everyone on the far end of a call was "Other". Turn on Tell remote speakers apart in Settings → Transcription and they become Speaker 1, Speaker 2 and so on, for up to four remote speakers. You can then assign each one to a person.

It runs on your machine, works with both Parakeet and Whisper, and needs a one-time model download. It works best in English and applies to new transcriptions. If it fails, the transcript falls back to Me and Other.

Run summaries on your own AI server

The built-in AI model takes several gigabytes of memory. Advanced users who already run llama.cpp, Ollama or LM Studio, locally or on another machine, can point MeetingDebrief at it in Settings → AI & Summaries: paste the address, add a token if it needs one, and click Test connection. Summaries, titles, chat and vocabulary suggestions then run on that server.

Your transcripts, notes and chat questions go to the server you choose, so pick one you trust. Audio doesn't: transcription always stays on your computer. The built-in model remains the default, and tokens are kept in your system's keychain.

I have been getting requests to support other AI servers and even online AI providers. If you run one that MeetingDebrief doesn't yet support and would like to see supported, please get in touch.

Integrations: let other apps react

Settings → Connections can now start a Zapier, Make or n8n flow, call Home Assistant, or run your own script (or a Shortcut on macOS) when a recording, transcript or summary finishes.

• Nothing is sent until you add an integration. By default it gets details only, such as ids, times and status. Sending meeting content is a separate choice for each integration.
• Limit it to some meetings by folder or tag, and pick which events it receives.
• Webhooks are signed using the Standard Webhooks spec, and brief outages are retried for up to 10 minutes. Anything older is dropped, so a laptop waking up the next morning doesn't replay yesterday.
• See what happened. Each integration has a Test button and a list of recent deliveries you can resend.

Local tools can also listen for events through GET /api/events on the API.

Choose where your files live

Settings → Storage lets you move recordings and the Markdown copies of your transcripts and summaries to a folder of your choice, such as an external drive or your Obsidian vault.

Recordings are always captured locally first and moved when they're finished, so an unplugged drive never costs you a meeting. Markdown files are only rewritten when their content changes, so edits you make in a notes app survive restarts.

Notes during the meeting, used in the summary

The notes field is now available while you record. Your notes go to summaries and chat as additional context, which helps with what mattered most to you and with names the transcript garbles. Custom templates can place them with a {notes} placeholder, and the API and MCP server can read them through a new get_notes tool.

A rebuilt Settings

Settings now has its own sidebar and shows one page at a time, grouped by what you're trying to do, with search across every setting. Some things have moved: models now live under Transcription and AI & Summaries, and feedback under About & Help.

Other improvements & fixes

• NEW: Dictations no longer include filler words and stutters like "um" and "uh". It's on by default, and you can turn it off with Remove filler words in Settings → Dictation.
• NEW: Insights shows the time dictation has saved you, compared with typing.
• On macOS, dictation only refuses to paste when the focused app holds Secure Input, and names the app. Failed pastes are harder to miss.
• Recordings now play as a centred mono mix, instead of your voice in one ear and the call in the other.
• Session titles follow the language of the summary.
• Failed sessions say why they failed, and tell "didn't start" apart from "transcription failed".
• If a saved system audio device disappears, recording uses the default output and tells you, instead of failing to start.
• Windows PCs with NVIDIA graphics download an updated Whisper GPU worker (over 200 MB) after updating. It brings Me and Other speaker labels and Vocabulary to GPU transcription with Whisper.
• Long meetings on Windows PCs with integrated graphics no longer run out of memory as easily. If they do, the summary retries with less memory, and otherwise shows a message you can act on.
• Progress toasts no longer get stuck when you leave a session during a job.
• The Move to menu stays visible for meetings near the bottom of the sidebar.
• MCP tools accept session ids from more clients and explain what's wrong when one is missing.

Version 5.1 lets you install models on networks that block downloads, send feedback from inside the app, and keep API and MCP searches to one folder or tag.

Highlights

Install models from files

If your network blocks our download servers, choose Can't download? Install from files in setup or Settings → Models. MeetingDebrief lists the exact files to download on another machine. Bring them over and click Import files, or drop them into the models folder. Each file is checked before it's installed. You can also skip setup; a banner shows what's still missing.

Send feedback from the app

Share an idea or report a problem from Settings → Feedback. You can attach a diagnostics bundle with recent logs, settings (secrets removed) and system info. Recordings, transcripts, summaries and notes are never included.

Folder and tag filters for the API and MCP

Search and session listings accept a folder filter (subfolders included) and a tags filter, and results now include each session's folder and tags. A tag filter that can't be used now returns an error instead of searching everything.

Other improvements & fixes

• NEW: Turn off "Me"/"Other" speaker labels in Settings → Audio if they often get lines wrong.
• Speaker labels no longer flip over a single stray word.
• Other apps' shortcuts no longer set off a modifier-only dictation shortcut like ⌃⌥.
• Diagnostics no longer report installed models as missing.

Version 5.0 adds dictation. Hold a shortcut anywhere on your machine, say what you want written, and the text is typed into whatever app you're in — an editor, a terminal, a chat box — then kept in its own history. It runs on the same local speech engine as your meetings, so nothing leaves your computer. Parakeet graduates from preview to the recommended engine, and the whole app gets a dark mode.

Highlights

Dictation: speak anywhere, text appears

Press and hold ⌃⌥D — or whichever shortcut you prefer — anywhere on your machine, say a sentence, and let go. The text is pasted into the app you were already typing in, and saved under the new Dictation section in the sidebar.

A small window shows what's happening while you speak: a level meter that moves with your voice, your words appearing as they're recognised, and how long you've been talking with a running word count. Press Esc to throw a recording away.

• Three ways to trigger it. Hold the shortcut like a walkie-talkie, press once to start and again to stop, or use both — a quick tap keeps recording hands-free, holding is push-to-talk.
• Your clipboard survives. Pasting has to go through the clipboard, so whatever you had copied is put back afterwards. Copy a log, dictate "take a look at this error", and the log is still there to paste underneath.
• Music pauses itself. Other audio is muted while you speak and restored exactly as it was — including the volume level, and including when you cancel. A system you had already muted is left alone.
• Your own words. Dictation runs through the same Vocabulary you taught MeetingDebrief, so product names and jargon come out spelled the way you mean them.
• A history you can search. Every dictation is kept with its audio for 30 days by default, so you can play one back or copy it again. The text stays until you delete it. Both the app's global search (⌘K) and the Dictation page search it, with dictations grouped separately from meetings so a result is never ambiguous.
• The tray can copy your last dictation, for when you dictated into the wrong window.

It's all in Settings → Dictation: the shortcut, how it starts and stops, whether it pastes for you, sounds, where the window appears, and how long audio is kept. On macOS, pasting for you needs Accessibility access; the panel asks for it and explains why.

Dictation needs the Parakeet speech model, which is fast enough to transcribe as you talk. A 15-second dictation is typed out within a fraction of a second of letting go.

Parakeet is now the recommended engine

Parakeet TDT v3 shipped as a preview in 4.0. It has lost the preview label and is now the model MeetingDebrief suggests during setup — 25 languages, punctuation and word timings, running locally on your machine's GPU where there is one.

Meetings and dictation now share one speech pipeline rather than each having their own, which makes the model behave better for both: it stays loaded between uses instead of being unloaded whenever a summary ran, so transcription and dictation start instantly. On machines with less than 16 GB of memory it still gives way when summaries or chat need the room. A dictation can also interrupt a long meeting transcription, take its turn, and hand it straight back — you never wait for a transcript to finish before you can dictate.

How long the model stays loaded is now set in Settings → Models, next to the model itself, and only appears once Parakeet is installed.

Dark mode

MeetingDebrief follows your system appearance, or you can pin it to light or dark in Settings → General. Every window is covered — the main window, the floating recorder, the dictation window, setup and updates.

Summaries know what day the meeting was

Ask a summary about "next quarter" and it used to guess, because it had no idea when the meeting happened. Summaries and chat are now told the meeting's date, so relative references — next quarter, end of the month, by Friday — resolve against the day you actually spoke rather than an invented one.

Choose which speakers to capture on Windows

Meeting audio on Windows used to record whatever the default output device was, which is wrong the moment your meeting app plays through a headset while Windows points at your monitor's speakers. Settings → Audio now lets you pick the output to capture. If a saved device disappears, MeetingDebrief says so instead of quietly recording only your microphone.

Other improvements & fixes

• NEW: The Dictation page walks you through your first dictation, shows your configured shortcut, and names whatever is stopping it — switched off, no speech model, or an unusable shortcut — with a button that fixes it.
• Pasting a dictation respects your keyboard layout, so ⌘V works on Dvorak and AZERTY rather than sending the wrong key.
• When an app blocks pasting — a password field, or an elevated window on Windows — the text stays on the clipboard and the dictation window tells you to paste it yourself.

Version 4.0 gives MeetingDebrief a second transcription engine, teaches it the words it keeps getting wrong, and gets out of your way while it works. Parakeet arrives as a preview alongside Whisper, Vocabulary replaces the old word-replacement list with something that actually learns, and long transcriptions and summaries now run in the background so the app stays usable while they do.

Highlights

A second transcription engine: Parakeet (Preview)

MeetingDebrief is no longer Whisper-only. Parakeet TDT v3 (Multilingual) — NVIDIA's 0.6B speech model — is now available as a second engine, covering 25 languages including English and Dutch, with punctuation, capitalization and word timestamps built in.

Download it from Settings → Models (~714 MB) and pick Use. It runs entirely on your machine, like everything else here: no cloud inference, no Python environment, no ONNX install, no NVIDIA GPU required. On Apple silicon it runs on the GPU via Metal. Live captions work too — Parakeet listens in buffered windows and revises what it heard as more audio arrives, so captions firm up as you talk.

Your existing setup is untouched. Whisper stays selected if that's what you're using, each recording remembers the model that produced it, and the final transcript flows into the same corrections, summaries, chat, search, export and Markdown mirror as before.

This is a preview, please let me know how it behaves.

Vocabulary: teach it your jargon once

Names, acronyms, product names and internal shorthand are exactly what local speech models mangle. The old Word Replacements list handled that after the fact. Vocabulary replaces it with something that works at three points at once: it primes Whisper before it listens, corrects the transcript as it's saved, and explains your terms to the model that writes summaries and answers chat questions — so "WHOOP" stops becoming "hoop", and the summary knows what it is.

Terms carry a meaning and any number of aliases, and the list grows as you work:

• From your edits — fix a word in a transcript and MeetingDebrief offers to remember it: "Always fix 'Kubernetis' → 'Kubernetes'?"
• From your transcripts — select a term in any transcript and add it straight from the floating chip.
• From the AI — Suggest terms reads a meeting and proposes the product names, acronyms and jargon it found.

Your existing word replacements move over automatically. Regex, substring and case-sensitive rules stay exactly as they were, now under Advanced corrections inside the Vocabulary panel.

Transcription and summarization run in the background

Starting a transcription or a summary no longer blocks the app. Both now run as background jobs with live progress and a Cancel button, so you can browse other meetings, take notes, or start a recording while a long transcript is still processing — and the progress follows you around the app rather than living in one view. Sessions can no longer get stuck in "summarizing" when something fails, and on Windows the worker processes no longer flash up console windows.

Force a language when detection gets it wrong

Whisper occasionally misidentifies similar languages — Finnish read as Swedish will derail an entire transcript. Re-transcribe is now a split button: the main half re-runs with auto-detection, and the attached caret opens a searchable language picker to force the right one. The detected language shows in the session header, and the sidebar gives each meeting a distinct icon for awaiting transcription, in progress, error, and done.

When the selected model won't honour a forced language, the picker now says so instead of silently ignoring the choice — Parakeet always auto-detects, and an English-only Whisper build is pinned to English.

Summaries tell you when they're out of date

Transcripts now carry a revision number, and every summary records the version it was written from. Edit a transcript or re-transcribe a meeting and its existing summaries are flagged "Transcript changed — regenerate summary" rather than quietly describing a transcript that no longer exists.

Other improvements & fixes

• NEW: The floating recorder gets an Open main window button, so you can jump back to the app without going through the tray menu.
• NEW: The local API can be exposed on your whole network instead of just this machine — off by default, in Settings → API. Useful for integrations running on another device; leave it off if you don't need it.
• NEW: Vocabulary is reachable from the local REST API, so connected tools can read and add terms.
• Vocabulary terms can be imported and exported as plain text, so a shared team glossary is a copy-paste away.
• FIXED: Summaries could hang before they even started on Windows machines that silently drop loopback connections — the first readiness check to the summarization engine waited forever, and the stuck job kept the "busy" flag set, so no further summary could run even after cancelling. Those checks are now bounded and cancellable.
• FIXED: License activation against an unreachable server hung indefinitely instead of reporting a network error. License requests now time out and fail fast.
• FIXED: Opening MeetingDebrief could start a second copy when the previous instance's handshake file was gone (macOS) or the running copy was elevated (Windows). Two copies shared one database and fought over the API port, which broke the MuteDeck integration. A lock file in the app data folder now settles it, and restarting after an update is no longer mistaken for a duplicate launch.
• FIXED: Live captions survive a window reload and recover if the transcription worker crashes, instead of going silent for the rest of the recording.
• FIXED: Cancelling a transcription no longer publishes a partial transcript or leaves the session in a stale state.

Version 3.3 focuses on the quality and reliability of your summaries: they now take in the whole meeting at once, always finish what they were writing, and make better use of your GPU. Transcription in languages other than English is faster, too.

Highlights

• IMPROVED: Summaries now see the whole meeting — MeetingDebrief sizes the model's working memory to your transcript, so very long meetings are summarized in a single pass whenever your hardware allows, instead of in stitched-together parts. Chatting with a meeting benefits from the same change: answers are grounded in the full conversation, even for hours-long recordings.

• IMPROVED: Long summaries finish cleanly — the model now gets all the writing room your hardware has available, and automatically picks up where it left off if it reaches its output limit. In the rare case where space truly runs out, the summary says so instead of stopping short.

• IMPROVED: Better GPU acceleration on Windows (NVIDIA) — summarization is now more dependable about running on the GPU, and the model engine checks and repairs its own installation on update — no reinstall or manual download needed. Each meeting also shows a small GPU/CPU icon, so you can see at a glance which one produced its summary.

• IMPROVED: Sturdier summarization — if the summarization engine runs into trouble mid-summary (typically on machines with little free memory), MeetingDebrief restarts it and retries automatically. On Macs, the model's memory use is now matched to your machine to avoid getting into that situation at all.

• IMPROVED: Faster transcription in other languages — the spoken language is detected once per recording rather than repeatedly, and tricky audio passages are retried less aggressively.

Other improvements & fixes

• Opening MeetingDebrief while it's already running brings the existing window to the front instead of starting the app twice.
• The local API binds exactly the port you configure — and lets you know if it's taken — so integrations that depend on a fixed port (like MuteDeck) keep working.
• Deactivating your license now asks for confirmation first.
• Summary edits in progress and search highlights no longer carry over when switching to another meeting.

Version 3.2 makes your meetings yours to edit and organize. Fix a stray word in a summary or transcript, jot notes as you go, tag and find meetings in seconds, and teach MeetingDebrief the names and jargon it keeps getting wrong.

Highlights

Edit your summaries and transcripts

Auto-generated text isn't always perfect. Now you can fix it. Hover any summary or transcript line and edit it in place — correct a misheard word, tidy a heading, or rewrite a sentence. Your edits are saved, picked up by search, and written through to the on-disk Markdown mirror, so the corrected version is the one you'll find later and the one your other tools see.

Notes for every meeting

Add your own notes to any recording — agenda, follow-ups, the thing you didn't want to forget. Notes are fully searchable alongside your transcripts and summaries, and they're included as context when you chat with a meeting, so the assistant answers with what you wrote, not just what was said.

Tag and find your meetings

Give meetings free-form tags — standup, client-acme, 1on1 - anything you want. Tags show up in the session header, and you can search for them with tag: filters: tag:standup, or combine them with and/or to narrow things down. Search a tag on its own to pull up whole meetings, or pair it with a phrase to scope your search to just those meetings.

Teach it your vocabulary

Names, acronyms, product names, and jargon are exactly what local speech models tend to mangle. The new word-replacement dictionary lets you fix that once and for all: add the misspelling and the correction, and MeetingDebrief applies it automatically. No more correcting "Kubernetes" or your coworker's name in every single meeting.

Other improvements & fixes

• FIXED: Recordings could finish with no transcript when a stretch of silence followed speech, leaving the speech detector to hand the local model an empty buffer and fail the whole recording. This is fixed.

• FIXED: A change in 3.0.0 made transcription noticeably faster, but it also introduced a regression: recordings could be cut off by a fixed time limit and finish with no transcript at all. Non-English speech takes the local model longer to process, so those recordings — along with any long meeting — were the most likely to hit the limit. This is now fixed.
• NEW: The sidebar is now a full folder tree, so you can create nested folders. Drag sessions and whole folders to rearrange them, and use the Move to menu to file a meeting away in a couple of clicks.

Version 3.0 lets you chat with your meetings, teaches MeetingDebrief to tell who's talking, and gives your recordings a serious audio-quality glow-up. Ask questions and get answers, know who said what, and enjoy cleaner sound in and sharper transcripts. It's also a lot quicker.

Highlights

Chat with your transcript

Long meeting? Just ask. The new chat panel lets you ask questions about any recording; "What were the action items?", "What did we decide about the budget?", "What did Sarah commit to?", and get a streamed answer drawn from the whole transcript, not just a keyword match. Answers cite the moments they come from: click a timestamp and MeetingDebrief jumps the audio and transcript right to that point. For long meetings, it automatically pulls in the most relevant parts so even hours-long recordings stay answerable. Open it from any meeting with Chat with this meeting (or ⌘L / Ctrl+L), and as with everything in MeetingDebrief, your conversation never leaves your computer.

Know who said what

Transcripts now label each segment by speaker. When you record both your microphone and system audio, MeetingDebrief separates Me from Other automatically, so a back-and-forth reads like a real conversation instead of one long block of text. Speaker labels show up in the history view and are included when you copy a transcript. Naming speakers is on the roadmap, but for now, "Other" will have to do.

Echo cancellation

No more hearing yourself twice. When the other side's audio leaks back into your microphone, MeetingDebrief now cancels that echo — during recording when possible, with an automatic offline cleanup pass as a fallback. The result is a clean mic channel and far fewer doubled-up words in your transcript.

Cleaner recordings, better transcripts

We added real-time noise suppression, a high-pass filter, and automatic gain control to the recording pipeline. Background hum, keyboard clatter, and uneven volume get smoothed out before transcription ever runs — which means more accurate text, especially in noisy rooms and quiet talkers alike.

No more phantom subtitles (or Hallucinations during silence)

Local speech models sometimes invent text during silence — stray "thank you for watching", "🎵 music 🎵 ", or the same phrase repeated on a loop. MeetingDebrief now recognizes and removes these known artifacts, so your transcript reflects what was actually said.

Bring your meetings to your AI tools

MeetingDebrief now ships a built-in MCP server, giving AI assistants like Claude Desktop, Claude Code, and Cursor read-only access to your transcripts and summaries. Ask your favorite agent to search across meetings, pull up a transcript, or summarize what you committed to last week — without copy-pasting anything. Settings → MCP gives you a ready-to-paste config with a Copy button to wire up a client in seconds. It runs on demand as a local sidecar, opens no network ports, and — like everything in MeetingDebrief — nothing leaves your machine.

Other improvements & fixes

• Rebuilt audio resampling to run a single high-quality pass at the device's native rate for cleaner, more faithful sound.
• Fixed a resampling issue that could truncate very short audio clips.
• Improved macOS system-audio (loopback) capture reliability, with better diagnostics when a device misbehaves.
• Added a full-text search endpoint to the local API, powering cross-meeting search for connected AI agents.

This is a big one! Version 2.0 brings powerful new ways to find, export, and understand your meeting data — plus major improvements to audio capture and GPU performance on Windows.

Highlights

• NEW: Audio File Import — Drop in an audio file (MP3, WAV, M4A, and more) and MeetingDebrief will convert, transcribe, and summarize it automatically. Great for processing recordings from other apps or devices.

• NEW: Session Mirror — Automatically sync your transcripts and summaries to disk as Markdown files. Perfect for backing up your data or integrating with tools like Obsidian or Notion. Click the Open Sessions button in Settings -> About to see the folder where these files are stored.

• NEW: Global Search — Instantly search across all your transcripts and summaries with Ctrl+F (Cmd+F on Mac). Full-text search so you can find that one thing someone said three weeks ago.

• NEW: Insights Dashboard — See how much time and money MeetingDebrief is saving you, based on your meeting history. A fun way to see the value of your summaries at a glance.

• IMPROVED: Processing speed on Windows — Both transcription and summarization are now GPU-accelerated on Windows if you have an NVIDIA CUDA-compatible GPU. This can make processing up to 10x faster! The processing has also been optimized for less powerful machines, so you should see improvements even without a GPU.

More new features

• Summary language override — force summaries into a specific language instead of relying on auto-detection
• Audio file retention setting — control how long recorded audio files are kept to manage disk space
• Performance priority setting — choose between faster processing or lower system impact during transcription and summarization
• GPU-accelerated summarization on Windows — NVIDIA CUDA support now extends to the summarization model, not just transcription
• Improved audio capture on macOS - use CoreAudio for better performance and reliability when recording system audio.

Improvements & fixes

• Dates and times now follow your OS locale settings
• Model files are validated at startup with a notification if anything is missing
• Concurrent transcriptions and summarizations are now prevented to avoid conflicts
• Fixed transcript loss caused by non-UTF-8 characters in audio
• Clearer setting names and descriptions throughout

• NEW: Real-time transcription — see your transcript appear live as you record
• NEW: Folder system for organizing recording sessions in the sidebar
• NEW: GPU-accelerated transcription on Windows to make it 10x faster
• NEW: New setting to control recording quality for better playback
• IMPROVED: Collapsible transcript area to save screen space
• IMPROVED: Automatically follows system default input device changes during recording
• FIX: Startup recovery for sessions stuck in a recording or summarizing state
• FIX: Transcription duration no longer inflated by system sleep time

• NEW: Add audio player with transcript sync for session playback, if recording is available
• NEW: Multi-language support with automatic language detection for transcripts and summaries (no need to do anything, just speak your language and it should work)
• NEW: Add a microphone test to check if the mic is working and see volume levels
• IMPROVED for macOS only: Transcribing and summarizing now try to use the GPU if available, for faster processing
• IMPROVED: Add logging for troubleshooting issues with transcription, summarization, and audio playback (settings now has an "Open logs" button that opens the log file in the system file explorer)