Voice Chat
Talk to the agent and hear it talk back. Voice and text share the same conversation — see Text Chat for the rest of the chat UI.
Finding a voice-enabled agent
Not every agent supports voice. Only agents with a speech model configured show the voice button in chat.
On Home:
- Click Filters and choose With voice support.
- Matching cards show a speech model badge (musical-note icon, e.g. Amazon Nova 2 Sonic).
You can also expand Already have an agent ID? at the top of Home, paste an ID, and press Start Conversation — useful when an admin shared an ID directly. See Home.
Switch to voice mode
In the chat page, use the input bar at the bottom. When the field is empty, a dark waveform button appears on the right — click it to enter voice mode.
Microphone permission
The first time you use voice on a given agent, the input bar is replaced by an in-app status bar:
Please enable microphone access to start voice chat.
Your browser also prompts for microphone permission — click Allow. Once permission is granted, the WebSocket connects and recording starts automatically.
If you denied permission by mistake: Browser settings → Site permissions → Microphone → Allow for this site. Refresh the page and click the waveform button again.
Voice controls
While voice mode is active, the input bar becomes a voice control bar:
| Control | Appearance | When visible | What it does |
|---|---|---|---|
| Mic / Unmute | Blue mic icon | While recording | Active mic; click to mute |
| Muted mic | Red mic icon | While muted | Click to unmute |
| Stop | Stop icon | While the assistant is speaking | Cuts agent audio immediately; your mic listens |
| End voice mode | X button | Any active voice session | Leaves voice mode and returns to text input |
| Restart | Restart icon | After session timeout only | Starts a new voice connection in the same conversation |
While muted, the bar also shows:
Microphone muted. Click Unmute to resume.
Muting only stops your mic — the agent can still speak. Prolonged silence (including while muted) can still trigger a session timeout.
How a voice turn works
- You speak — Voice Activity Detection (VAD) detects when you start and stop talking.
- The agent thinks — a short pause while the response is prepared.
- The agent replies out loud — audio streams in real time.
- Control returns to you — the mic is active again automatically.
Repeat until you end the session or it times out.
Interrupting the agent (barge-in)
You do not need to wait for the agent to finish:
- Click Stop while the assistant is speaking — audio cuts off immediately and the mic listens.
- Start speaking — VAD detects your voice after a brief delay and barges in automatically.
Use Stop when you need an instant cutoff; just speak when a natural overlap is fine. This is the fastest way to redirect when the agent goes off track.
Session timeout and restart
Voice sessions close after a period of silence to free server resources. When that happens, the controls collapse to a Restart button only — click it to open a new voice connection in the same conversation. Message history is preserved.
Mixing voice and text
Voice and text are on the same conversation thread. Click End voice mode to type, or click the waveform button again later. Quick voice questions plus occasional typed corrections work well.
Connectors and voice
If required connectors are not authorized, a banner appears above the input:
Connectors not connected. Some tools may not work, but the agent will. Connect now
Voice chat still works — the agent will talk to you. Tools that depend on connectors (for example smart-home controls) will not run until you click Connect now and finish authentication. See Connectors and Text Chat — Connecting a service mid-chat.
Audio quality
The dashboard uses 8 kHz PCM audio in both directions. That keeps bandwidth low while staying clear enough for conversation, including on slower networks.
The speech model on the agent card (e.g. Amazon Nova 2 Sonic) sets voice quality and language capability; different agents may use different models.
Tips
- Use a quiet room. Background noise hurts recognition.
- A wired or close-by mic helps. Laptop mics are fine; distant Bluetooth headsets are not.
- Speak naturally. No need to over-enunciate.
- Use barge-in freely. Faster than waiting through a wrong answer.
- Clear the input before switching to voice. Unsent typed text is discarded.
Troubleshooting
| Problem | Try |
|---|---|
| In-app banner: "Please enable microphone access" | Allow the mic in browser site settings, then click the waveform button again. |
| Mic permission was denied | Browser settings → Site permissions → Microphone → Allow for this site, then refresh. |
| No voice button | This agent has no speech model. On Home, use Filters → With voice support. |
| Voice button missing after allowing the mic | Hard refresh (Ctrl+Shift+R / Cmd+Shift+R) and try again. |
| Choppy audio | Check your network; try wired instead of Wi-Fi if possible. |
| Agent didn't hear me | Confirm the mic is not muted (red icon); check the input device in OS sound settings. |
| No sound from the agent | Check OS speaker output; ensure the browser tab is not muted. |
| Agent stopped responding | Refresh the page — voice WebSockets sometimes need a reconnect. History is preserved. |
| Filters show no voice agents | No voice-enabled agents in your current view; ask your admin or try All. |