Skip to main content

Voice Production Deployment

Checklist and architecture for deploying voice agents in production.

TLS / WSS Configuration​

Voice WebSocket connections must use WSS (WebSocket Secure) in production. Configure TLS at the reverse proxy level:

# nginx example
location /voice/ws {
    proxy_pass http://localhost:3777;
    proxy_http_version 1.1;
    proxy_set_header Upgrade $http_upgrade;
    proxy_set_header Connection "upgrade";
    proxy_read_timeout 86400;
}

For telephony webhooks (Twilio/Telnyx/Plivo), the webhook URL must be HTTPS. Use a reverse proxy or ngrok for development.

Telephony Webhook Security​

Always verify webhook signatures in production:

{
  "telephony": {
    "provider": "twilio",
    "verifySignatures": true,
    "authToken": "your-auth-token"
  }
}

Without signature verification, anyone can send fake call events to your webhook endpoint.

Monitoring​

Key metrics to track:

MetricTargetAlert Threshold
STT latency (P95)<500ms>1000ms
TTS latency (P95)<300ms>800ms
End-to-end turn latency<2s>4s
Barge-in rate<30%>50% (agent talking too long)
STT error rate<2%>5%
Call drop rate<1%>3%
Concurrent callsvaries>80% capacity

Capacity Planning​

Each active voice call consumes:

  • ~1 WebSocket connection
  • ~2-5 MB RAM (audio buffers, session state)
  • ~64 kbps bandwidth per direction (mu-law telephony) or ~256 kbps (raw PCM WebSocket)

Rule of thumb: a single 2-core server handles 50-100 concurrent voice calls comfortably.

Call Logging​

Enable call logging for debugging and compliance:

{
  "telephony": {
    "logCalls": true,
    "logDir": "./call-logs",
    "retentionDays": 90
  }
}

Logs include: call metadata (duration, participants, DTMF), full transcript with speaker labels and timestamps, and agent decisions.

Multi-Region Architecture​

For low-latency voice, deploy close to your users:

  • US East + US West for North America
  • EU West for Europe
  • AP Southeast for Asia-Pacific

Each region runs its own voice server connecting to the nearest STT/TTS provider endpoints. The LLM backend can be centralized (latency is less sensitive for the reasoning step).

Graceful Degradation​

Configure fallback behavior when providers fail:

{
  "voice": {
    "stt": {
      "provider": "deepgram",
      "fallback": "whisper-local"
    },
    "tts": {
      "provider": "elevenlabs",
      "fallback": "openai"
    }
  }
}

If the primary provider times out (>5s) or returns an error, the pipeline automatically switches to the fallback provider for the remainder of the call.