Docs
Short guides to everything in EchoDiar, on Mac and Windows.
Install
macOS
- Download the signed and notarized DMG from Releases, drag EchoDiar to Applications.
- Needs macOS 15+ on Apple Silicon. Apple Speech and Apple Foundation Models need macOS 26.
- EchoDiar lives in the menu bar. Its Dock icon shows while one of its windows is open.
- Updates itself: menu bar icon > Check for Updates…
Windows
- Get EchoDiar from the Microsoft Store: the Store installs it in one click. There is no certificate to import.
- Needs Windows 11 on x64 or Arm64.
- Updates through the Microsoft Store: the title bar shows Update when a new version is out, and tray icon > Check for updates… checks now.
- Installed EchoDiar earlier with its certificate (EchoDiar.cer)? Move to the Microsoft Store.
Moving to the Microsoft Store
Earlier versions of EchoDiar for Windows were installed with EchoDiar's own certificate (EchoDiar.cer and EchoDiar.appinstaller). That copy is not signed by a public certificate authority, which is why Windows warned during the install and why Smart App Control may block it. The Microsoft Store copy is signed by Microsoft. To move, once per PC:
- Uninstall EchoDiar in Settings > Apps > Installed apps. Your recordings and settings in
%LOCALAPPDATA%\EchoDiarstay, and so do your API keys in Credential Manager. - Install EchoDiar from the Microsoft Store. It opens your recordings as before.
- Remove EchoDiar's certificate: Start, search "computer certificates", open Manage computer certificates (click Yes when Windows asks for permission), go to Trusted People > Certificates, right-click the certificate issued to Anh-Thi Dinh, click Delete, then Yes.
Permissions
- Microphone: records your voice. On Windows, if it is off, turn on EchoDiar in Windows Settings > Privacy & security > Microphone.
- System Audio Recording (on Mac): records the meeting's audio. The first-run onboarding tests it and opens the right Privacy & Security pane. On Windows, the onboarding's Test System Audio checks that EchoDiar hears it.
- Accessibility (optional, on Mac): only to keep the floating indicator on the meeting window. Windows needs no permission for that.
- Calendars (optional): reads your meetings to offer recording when they start and to name the recording, and adds the events you create to an "EchoDiar" calendar. EchoDiar writes to no other calendar. On Mac, it reads macOS Calendar, and the "EchoDiar" calendar is On My Mac when available, otherwise in your default account, e.g. iCloud, which syncs it. On Windows, click Grant Access in Settings > Calendar; if access is off, turn on EchoDiar in Windows Settings > Privacy & security > Calendar.
- Tell the people you record. Onboarding and Settings > General have a Copy consent message button.
Recording
- Record one app's audio or all system audio, plus your mic. Mic and system audio go to separate channels of one file.
- When an app's audio cannot be isolated, EchoDiar records all system audio and says so.
- On Mac, start or stop with Cmd+R in the app, or Cmd+Option+R from any app. Pause with Cmd+Shift+P in the app, or Cmd+Option+P from any app.
- On Windows, start or stop with Ctrl+R in the app, or Ctrl+Alt+R from any app. Pause with Ctrl+Shift+P in the app, or Ctrl+Alt+P from any app.
- Live level meters while recording, and live captions with Transcribe while recording. The menu bar icon on Mac shows idle, recording, paused and processing. On Windows, the notification area icon shows a red dot while recording and an orange badge while paused; a left click opens EchoDiar.
- Pick the microphone, and what happens when it disconnects, in Settings > Recording > Microphone.
- Each recording is compressed to AAC (64 kbps by default) once it is transcribed. A recording cut short by a crash is recovered and transcribed at next launch.
Meeting detection
- Offers to record when a meeting app starts using the mic: Zoom, Teams, Slack, Webex, Discord, Skype, FaceTime on Mac, and browsers for web calls such as Meet.
- Per app: Ask, Always record or Ignore (browsers: Ask or Ignore).
- A floating indicator shows while recording. Drag it, hide it while EchoDiar is in front, or attach it to the meeting window.
- Auto-stop: when the meeting app quits, the call goes quiet or the screen locks, the indicator asks before stopping, with a 15 s countdown.
Calendar
- On Mac, works with any account added to macOS Calendar: iCloud, Google, Outlook/Exchange, CalDAV. On Windows, reads the calendars Windows shares with apps. Set it up in Settings > Calendar, with the optional Calendars permission.
- When a meeting starts: Off, Ask (the default: the floating indicator offers Record, Join or Not now), Wait for the call (records the meeting app as soon as the call starts) or Record at start (records all system audio and the mic at the start time).
- The automatic modes start without asking only for events you organize, accepted or answered Maybe to, and events without invitees in the EchoDiar calendar (or, on Mac, a calendar On My Mac). Other events ask.
- Choose the calendars to watch, and keep only events with a meeting link (Zoom, Google Meet, Microsoft Teams, Webex).
- Stop when the event ends, plus a few minutes (10 by default). Once the recording has also been silent for 1 min, the indicator asks before stopping, with a 15 s countdown, so a meeting that runs over keeps recording.
- A recording started from an event is named after it.
- The Calendar page (toolbar button next to Settings) lists the meetings of the next 7 days (updated automatically) with Join, Record and Skip. The menu bar menu on Mac, and the notification area icon's menu on Windows, list upcoming meetings.
- New Event… on the Calendar page creates a meeting: title, date, start and end, and an optional meeting link (https). It goes to an "EchoDiar" calendar: on Mac, On My Mac when available, otherwise your default account, e.g. iCloud, which syncs those events like any other; on Windows, EchoDiar's own calendar, which no other app can add events to.
- Repeat: None, Daily, Weekdays, Weekly or Monthly, with End Repeat never or on a date.
- Edit… on events in the EchoDiar calendar changes them, and Delete Event… is in the editor. For a repeating event, choose This Event Only or This and Future Events; when Repeat or End Repeat changed, Save offers only This and Future Events.
- Events in the EchoDiar calendar count even without a meeting link, and follow the same When a meeting starts mode. On Mac, don't share the EchoDiar calendar: anyone who can add events to it could start a recording.
- Works while EchoDiar runs, so turn on Launch at login in Settings > General.
Transcription
- Starts on its own: while you record with Transcribe while recording, otherwise after Stop. Turn it off with Transcribe automatically in Settings > Recording > After recording.
- Engines on Mac: WhisperKit, Parakeet, Qwen3-ASR, and Apple Speech on macOS 26.
- Engines on Windows: Whisper (whisper.cpp: Large v3 Turbo by default, Small, Base, Large v3, and PhoWhisper Medium, Vietnamese only) and Parakeet TDT 0.6B v3.
- Models are downloaded only when you click Download in Settings > AI Providers > Local AI. On Mac, macOS installs Apple Speech language assets the first time you use a language.
- Speakers: your mic is Me. Turn on Tell speakers apart in Settings > Recording > After recording (or Settings > Transcription) to label each voice in the meeting audio: Speaker 1, Speaker 2… Set the number of other speakers there as a hint. Off (the default), the meeting audio is one speaker, Speaker 1, and cloud transcription is asked for no speaker labels.
- Transcribe while recording (on by default, Whisper and Parakeet): transcribes during the recording and shows live captions, so the transcript is ready seconds after Stop. Uses more CPU and memory while recording. Turn it off in Settings > Transcription. See How it works. With a cloud engine, see Cloud transcription.
- Re-transcribe a conversation with another engine or language.
- On Mac, can wait while the Mac is hot or in Low Power Mode: what is left after Stop is done once the Mac cools down or leaves Low Power Mode.
- On Windows, can wait while energy saver is on, with Wait while energy saver is on (on by default) in Settings > Recording > After recording: what is left after Stop is done once energy saver is off.
How it works
- With Transcribe while recording on (the default), the recording is transcribed while it happens. Each channel, your mic and the meeting audio, is cut into chunks at pauses in speech, at most 12 s; with no pause, the cut falls on the quietest moment. Each chunk becomes one line of the live transcript. While a chunk is still open, its text so far shows in grey and is refreshed every few seconds; the chunk's line replaces it. When decoding falls behind, on a slower computer, the status says how many seconds behind it is.
- Chunks wait in a queue that keeps up to about 10 minutes of audio, and one speech model, loaded once for the recording, transcribes them. Its lines are the live captions, Me and Others, and they become the saved transcript. Whisper and Parakeet only.
- Nothing is lost: chunks beyond the queue and chunks that fail become gaps, transcribed from the recording after Stop. On Mac, chunks held while the Mac is hot or in Low Power Mode (with Wait when the Mac is hot or in Low Power Mode on) stay in the queue and are transcribed once the Mac recovers; chunks still held at Stop become gaps too.
+-----------+ +------------------------+ +-------------------------+
| Recording |-->| channel writer, 16 kHz |-->| audio.caf |
| mic + | +------------------------+ | ch0 = mic, ch1 = system |
| system | | +-------------------------+
+-----------+ v
+-----------------------------+
| speech chunker, per channel |
| ends on a pause, 12 s max |
+-----------------------------+
|
v
+-----------------------------+
| chunk queue, 10 min max |--> queue full, still held
+-----------------------------+ at Stop (Mac hot), or a
| chunk failed: gap,
v transcribed from
+-----------------------------+ audio.caf after Stop
| speech model, loaded once |
| Whisper or Parakeet |
+-----------------------------+
|
v
segments --> live captions (Me, Others)
- At Stop, the same model transcribes the last chunk and what is left in the queue, then fills the gaps. The transcript is written before the audio is compressed.
- Speakers: with Tell speakers apart off (the default), the meeting audio is Speaker 1 and the transcript is ready about 1 s after Stop (measured on a Mac): 0.8 s for a 15-minute meeting, against about 95 s when the whole recording is transcribed after Stop. On, the speaker model runs after Stop and labels Speaker 1, Speaker 2…: about 10 s for the same meeting.
- The transcript is made from the uncompressed audio, which is slightly more accurate: 0.0 % word error rate on the same meeting, against about 1 % from the compressed file.
Stop | v +------------------------------------------------+ | last chunk + queue, with the same model | +------------------------------------------------+ | v +------------------------------------------------+ | gaps, transcribed from audio.caf | +------------------------------------------------+ | v +------------------------------------------------+ | speakers, only if "Tell speakers apart" is on | +------------------------------------------------+ | v +------------------------------------------------+ | merge: mic = Me, system = Speaker 1 | | (Speaker 1, 2, ... with speakers on) | +------------------------------------------------+ | v +------------------------------------------------+ | transcript.json |<-- transcript ready +------------------------------------------------+ | v +------------------------------------------------+ | compress to audio.m4a, audio.caf deleted | | (moved to the Trash, or the Recycle Bin on | | Windows, not compressed, with "Delete audio | | after transcription") | +------------------------------------------------+
- Extra cost while recording, on top of recording alone, measured on a Mac: about 190 MB more memory for EchoDiar and about 20 s more CPU time per 15 minutes, about 2 % more of one core (debug build, averaged over a 15-minute meeting, MacBook Pro, M4 Pro). In a macOS virtual machine, which has no Neural Engine, the transcript is ready 3.1 s after Stop instead of 282 s.
- While a recording stays paused longer than Free unused models after, its model is freed; it loads again when you resume.
- Local AI (summaries, chat and semantic search on this computer) waits until Stop while a recording holds the speech model.
- Other cases take the slower path:
Fallbacks
|
+-- "Transcribe automatically" off
| --> compress only
|
+-- "Transcribe while recording" off
| --> whole recording after Stop, still before compression
|
+-- cloud engine, its destination already confirmed
| --> transcribe from audio.caf, then compress
|
+-- cloud engine or cloud speaker model waiting for your OK
| --> compress, ask, then upload
|
+-- Qwen3-ASR or Apple Speech (Mac)
| --> whole recording after Stop
|
+-- engine changed, a language the engine cannot transcribe,
| or the model failed to load
| --> whole recording after Stop
| (switching between supported languages keeps the fast
| path: each chunk carries its own language)
|
+-- Mac hot or in Low Power Mode, with "Wait when the Mac
| is hot or in Low Power Mode" on (on Windows: energy
| saver on, with "Wait while energy saver is on" on)
| |
| +-- no gaps, speakers off
| | --> transcript built while recording saved at once,
| | then compressed
| |
| +-- gaps from chunks held while hot, speakers on,
| or no transcript built while recording
| --> compress first; gaps, speakers and the transcript
| wait until the Mac cools down or leaves Low
| Power Mode (on Windows: until energy saver
| is off)
|
+-- crash
| --> the recovered audio.caf is compressed, then the
| whole file is transcribed at next launch
|
+-- re-transcribe
--> whole file with the chosen engine
Cloud transcription
- Optional: OpenAI, Groq, Mistral, Gemini, xAI, ElevenLabs, Deepgram, AssemblyAI, Speechmatics, Soniox, Gladia or a custom OpenAI-compatible endpoint.
- Save each key in Settings > AI Providers > Cloud, with Test and a link to get a key; custom endpoints are set up there too. A provider is available once its key is saved. Pick the model in Settings > Transcription: a checkmark marks each provider with a saved key, and API key opens its card in AI Providers > Cloud.
- The first run for each provider and address asks before any audio leaves your computer. A changed URL asks again.
- The mic and system channels are uploaded, and billed, separately. Speakers keep their labels when a long recording is uploaded in parts.
- Live transcript while recording (with Transcribe automatically and Transcribe while recording on), in the While recording section of Settings > Transcription. Live with a local model shows it with a model on your computer, free; it is display only, and the saved transcript still comes from the cloud provider after Stop.
- Live with … realtime streams to the default provider instead: ElevenLabs, Deepgram, AssemblyAI, OpenAI, xAI, Speechmatics, Soniox or Gladia (Mistral, Gemini, Groq and custom endpoints have no realtime). It costs money: two live connections, mic and system audio, for the whole recording, on top of the normal transcription after Stop. It asks first for a host you have not confirmed yet, naming it.
- Use the realtime result as the transcript skips the upload after Stop: only the parts missed while disconnected, and around pauses, are sent then. Realtime models can be less accurate.
- API keys are stored in the Keychain on Mac, in Credential Manager on Windows.
Transcript & playback
- Click a line to hear it. Space plays or pauses, Esc goes back.
- Edit the title, rename speakers, copy the transcript or reveal the folder in Finder on Mac, File Explorer on Windows.
AI summaries & chat
- Open the AI panel with Cmd+I on Mac, Ctrl+I on Windows. Generate a summary or ask questions about a conversation.
- Answers cite
[mm:ss]timestamps. Click one to jump to that moment in the transcript and recording. - Response language, for summaries and answers: automatic (the transcript's language for summaries, the question's for answers), English, French or Vietnamese.
- On-device: on Mac, MLX models, and Apple Foundation Models on macOS 26. On Windows, GGUF models (Qwen3.5 2B, 4B or 9B, Gemma 4 E2B or E4B), and Phi Silica on Copilot+ PCs.
- Local servers: LM Studio, Ollama, Jan, llama.cpp, vLLM, plus Foundry Local on Windows. Scan finds the ones running on this computer.
- Cloud APIs: OpenAI, Anthropic, xAI, Gemini, Mistral, Azure OpenAI, OpenRouter, Groq, DeepSeek or a custom endpoint. Keys go in Settings > AI Providers > Cloud, stored in the Keychain on Mac, in Credential Manager on Windows.
- Subscription CLIs: Claude Code, Codex, OpenCode or Gemini CLI, with your own login.
- Copy as prompt copies the transcript and your question, ready to paste into any chat app. On Windows, Ctrl+Shift+C does it from the Summary or Chat tab.
Library & search
- Conversations grouped by day, with their size on disk.
- Rename, reveal in Finder (Explorer on Windows) or move to the Trash (the Recycle Bin on Windows).
- Search titles and transcripts with Cmd+F on Mac, Ctrl+F on Windows.
Storage
- Data lives in
~/Library/Application Support/EchoDiaron Mac,%LOCALAPPDATA%\EchoDiaron Windows, with the same layout. Settings > Storage shows the space used. - Auto-delete: once a day, moves transcribed conversations older than an age you set to the Trash (the Recycle Bin on Windows), whole or audio only.
- The uncompressed audio (
audio.caf) is kept until the transcript is written, then compressed toaudio.m4aand deleted. If transcription can't start right away, the audio is compressed first. - Optionally delete the audio once a transcription succeeds: with Delete audio after transcription, a new recording's uncompressed audio is moved to the Trash (the Recycle Bin on Windows) instead of being compressed. The transcript, summary and chat stay.
- Choose the compression bitrate, or keep the uncompressed audio too.
Settings & shortcuts
- Open with Cmd+, on Mac, Ctrl+, on Windows. Sections: General, Recording, Calendar, Transcription, AI Providers, AI Settings, Storage, Shortcuts, About, Acknowledgements.
- Change the global record and pause shortcuts, and view the app's shortcuts, in Settings > Shortcuts.
- Launch at login and the floating indicator are in Settings > General.
- Default language, Remember last record language (off by default: keeps the language picked in the Record button's menu for the next recordings), Transcribe while recording and Free unused models after are in Settings > Transcription. Tell speakers apart is in both Settings > Recording > After recording and Settings > Transcription.