Jot 0.4.0 — live transcription

Signed, notarized and stapled. Requires macOS 14+ and your own Gemini API key.

Live transcription (experimental)

Settings → Dictation → Experimental → Live transcription. Off by default.

Audio streams to gemini-3.5-transcribe-live over a WebSocket while you speak, so words appear in the pill as you talk and the transcript is ready when you release the key instead of starting an upload then.

When smart transcription cleans the text up, the pill shows the edit: fillers and self-corrections are struck out, then the sentence closes up around them.

Turning it on cannot cost you words. The recording is still written to disk exactly as before, and a live transcript is only used when the session finished cleanly and the bytes the socket accepted reconcile against the audio on disk. Anything else — dropped connection, truncated stream, no transcript in time — falls back to uploading the recording. You would notice latency, never loss.

Settings reports how often it actually worked, so you can judge it rather than guess. It also stops trying after three consecutive failures, since a broken live path is slower than having the feature off; one success clears that.

Also in this release

• Dictionary terms bias the recogniser directly, and the same replacement rules run over live text, so switching modes does not change how names are spelled
• The API key goes in a header on the live path too, never a URL query string
• Fixed: long sentences drawing outside the pill

Full changelog: v0.2.0...v0.4.0

Jot 0.2.0

Signed, notarized and stapled — installs without a Gatekeeper prompt.

• Relicensed to Apache 2.0 (was MIT)
• Native smart transcription with Dictionary terms sent as recogniser vocabulary
• Secure-input refusal now names the app holding it and how to clear it
• "Start at login" toggle no longer goes stale against macOS

Requires macOS 14+ and your own Gemini API key.