Duplex with Automatic Muting (Interview-Style)

This guide describes how to use duplex mode with automatic mic muting during playback. Use it when you want:

• No push-to-talk button (hands-free like duplex)
• No accidental interrupts from ambient noise (coughs, other people talking) while the assistant is speaking
• Mic automatically on when the assistant finishes, off while they are speaking

Same backend as standard duplex; only the frontend capture and barge-in behavior change.

1. When to use

• Interview agents: interviewer agent asks a question, then the user answers without touching the keyboard.
• Presentations / guided flows: assistant speaks, then user responds; you do not want background noise to cut off the assistant.

Standard duplex is better when you want true barge-in (user can interrupt the assistant at any time).

2. Core idea

• Backend: unchanged. Use the same duplex session, events, and duplex_state_changed status.
• Frontend:
  • While status is speaking or awaiting_playback_drain: mute — do not send any duplex_audio_chunk and do not send duplex_interrupt. Optionally keep the recorder running but discard chunks, or pause capture entirely.
  • When status becomes listening: unmute — resume sending chunks and normal silence-based end-turn.

So: duplex with automatic muting = duplex + “mic off during playback, mic on when listening.”

3. Status-driven muting

Drive muting from the server status you already receive via duplex_state_changed:

When status transitions from speaking or awaiting_playback_drain to listening, unmute and start sending chunks again.

4. Implementation checklist

1. Reuse duplex setup
   Same connection, duplex_audio_chunk, duplex_end_turn, duplex_playback_drained, and playback pipeline as in duplex_frontend_setup.md.

2. Gate capture on status

   • If status is speaking or awaiting_playback_drain: do not send any duplex_audio_chunk (and do not call duplex_interrupt).
   • You can either stop the recorder while muted or keep it running and drop chunks; stopping avoids unnecessary CPU.

3. Do not implement barge-in

   • Ignore user speech for interrupt purposes while status is not listening.
   • Do not call duplex_interrupt in this mode.

4. Playback unchanged

   • Decode and play assistant audio as in duplex.
   • When playback is done (and when you force-stop on disconnect), send duplex_playback_drained so the backend can move to listening.

5. Unmute only on listening

   • When duplex_state_changed reports listening, (re)start capture and allow chunks and end-turn again.

5. Optional: thinking

You can also mute during thinking so no stray audio is sent before the assistant speaks. If you do, unmute only when status is listening.

6. Reference

• Duplex frontend contract and events: duplex_frontend_setup.md
• Backend duplex lifecycle: ../voice_modes/duplex_setup.md