Voice interaction

Voice chat on Amethyst

1. Hold the mic. The pill’s Voice option opens a compact panel: Hold to talk. The microphone permission is requested on your first press — never before, never in the background.
2. Recording shows. A red pulse and “Listening…” are visible for as long as audio is being captured, capped at 60 seconds. Tap again to stop, ✕ to discard.
3. Transcription. The audio blob goes to the transcription gateway, comes back as text, and the raw audio is dropped — it is never written to disk, the database, or a log.
4. The transcript becomes your question. It flows through the exact same send path as typed text: same session conversation, same context limits, same streaming reply you can stop.
5. The reply is spoken. Sentence-sized chunks stream to the speech gateway and play back in order, gapless. A Stop control ends playback and pending chunks.

Why it feels fast

Speech synthesis runs on a primary engine that accepts a whole reply in one request (typically under three seconds for a full answer), with a fallback engine tried sequentially only if the primary fails. Silence is trimmed from the audio, and the client chunks the next sentence while the current one plays — so there is no dead air between sentences.

Honest limits

• There is no wake word and no background listening — this is push-to-talk by design.
• Voice features require a signed-in session; signed out, the app says so instead of faking it.
• Speech output is English-first today, and spoken text is capped (6,000 characters per reply) as an anti-abuse bound — the written reply is never truncated.
• Windows microphone privacy settings apply: if the mic toggle is off, dictation explains exactly which switch to flip.

Raw audio retention: none. Only the transcript and the reply text are saved, exactly like a typed exchange.