Voice
Everything spoken becomes text on your machine. Smritsavant bundles Whisper and runs it in-process — no audio is ever uploaded, and voice works even with the AI chat provider unconfigured or off. (On Windows and Linux, read ⌘ as Ctrl.)
Three ways in
| Entry point | What it does |
|---|---|
| ⌘⇧V Dictation | Transcribes into the open editor as you speak |
| ⌘⇧R Lecture capture | Long-form recording session, transcribed into a note with the audio attached |
| Drag & drop | Drop .mp3 .wav .m4a .webm .ogg .flac onto the editor to import + transcribe |
The transcription model (~1 GB) downloads once, when you first use a voice feature.
Languages
Whisper auto-detects the spoken language. If it guesses wrong, each transcribed note has a retry chip with a searchable language picker — re-run the transcription in the language you actually spoke. For low-resource languages, a local translation fallback can produce an English transcript when direct transcription quality is poor.
Transcripts are first-class
The audio block keeps the recording and its transcript together, and the transcript feeds the same pipeline as typed text:
- Spoken notes are semantically embedded and auto-linked like everything else.
- They match in both full-text and semantic search — “that thing I said in Tuesday’s lecture” is findable.
- Exports include transcripts alongside the audio files — see Import & export.
Practical notes
- Transcription runs in the background queue; you can keep writing while a long lecture processes. The processing chip in the top bar shows progress.
- Recording uses the system microphone permission; Smritsavant asks the first time.
- On mobile, the voice model download is Wi-Fi-gated by default like all model downloads — see Mobile.