Voice messages, transcription and dictation (on-device)
Send voice messages in WidgetStore Messages, transcribe them on your own device, and dictate into chats and Memos. Your voice is never uploaded.
Messages supports voice messages: hold to talk, release to send, even when the other person is offline. Voice messages you receive can be transcribed on your own device, and you can dictate into the composer and Memos. Recognition always runs on local offline models; your voice is never uploaded to any server.
Send a voice message
- At the bottom of a chat, click the sound-wave icon (or press Alt+V). The text box turns into Hold to talk.
- Phone: hold to talk and release to send. Slide your finger up before releasing to cancel.
- Computer: hold and release to send, or click once to start recording, then click Send (or press Alt+V again). Esc cancels.
- Up to 60 seconds. In the last 10 seconds you'll see a countdown, and it sends automatically when time is up. Click the keyboard icon to go back to typing.
Voice messages are end-to-end encrypted like text and photos. If the other person is offline they wait in the encrypted inbox (up to 7 days, deleted on delivery). They work in group chats too. Audio is Opus (AAC on Safari), about 180 KB per minute.
Listen
- Tap play to play or pause. Click or drag on the waveform to jump anywhere. While playing, click 1× to switch to 1.5× or 2× (remembered).
- Unplayed voice messages have a red dot. When one finishes, the next unplayed voice message plays automatically, even with text in between.
Transcribe
- Click Transcribe under a voice message. The first time, you're asked to download a local model (the size is shown; downloaded once). Then you can turn on Transcribe voice messages automatically in Messages settings → Voice and transcription.
- The text appears under the message; you can hide it, copy it or transcribe again. It's kept on this device only and isn't sent to anyone.
- Senders can turn on Add text to voice messages I send: the message is transcribed on your device before sending and the text is encrypted together with the audio.
Which model is used:
| Device | How | Models |
|---|---|---|
| Desktop app (Apple silicon Mac, Linux) | whisper.cpp on your computer (Metal-accelerated on Mac) | tiny ~32 MB, base ~60 MB (recommended), small ~190 MB |
| Web, desktop app on Windows / Intel Mac, phones | Local model in the browser (WebAssembly) | tiny ~43 MB, base ~80 MB |
All are multilingual Whisper models that understand Chinese, English, Japanese and Korean. Models are downloaded from Hugging Face (a mirror is used automatically where needed), stored on your device, and can be deleted under Voice and transcription. Speed depends on your device; usually a few seconds per message.
Dictation
Click the microphone in the chat composer or in Memos (in chats you can also press Alt+D). What you say shows up live; click again to finish and the text is inserted at the cursor so you can edit it before sending.
- Local models are used first. Without one, you can download a local model or use the browser's built-in speech recognition — built-in recognition may send your voice to the browser maker's servers (such as Google or Apple), and you're told so the first time.
- Up to 2 minutes per take.
FAQ
- "No microphone access": allow the microphone from the icon left of the address bar or in system settings. The desktop app shows the system prompt the first time you record.
- "This device can't transcribe yet": very old browsers (no WebAssembly or Cache Storage) can only play voice messages. Sending and receiving still work.
- Does it cost anything? No. Recognition runs entirely on your device, costs us nothing and uses no Sparks.
- "Couldn't make that out": recognition uses your app language (if people speak another language, change it under Messages settings → Voice and transcription → Speech language). When the result looks garbled, keeps repeating or is in the wrong language, we don't show it — try again somewhere quieter or speak a little slower. In the browser, Chinese, Japanese and Korean use the standard (base) model.
Last updated: Nov 18, 2026