Voice Adapter Capability Matrix
Every voice adapter in Scenario declares its capabilities via a frozen
AdapterCapabilities
dataclass. The table below is auto-generated from those declarations and kept
in sync by a CI gate — if you change an adapter's capabilities, regenerate
with:
cd python
uv run python scripts/gen_capability_matrix.pyCapabilities
| Adapter | streaming_transcripts | native_vad | dtmf | interruption | input_formats | output_formats |
|---|---|---|---|---|---|---|
| ComposableVoice | ✅ | ❌ | ❌ | ❌ | pcm16/24000 | pcm16/24000 |
| ElevenLabs | ✅ | ✅ | ❌ | ❌ | pcm16/24000 | pcm16/24000 |
| GeminiLive | ✅ | ✅ | ❌ | ✅ | pcm16/16000 | pcm16/24000 |
| LiveKit | ✅ | ✅ | ❌ | ❌ | pcm16/48000 | pcm16/48000 |
| OpenAIRealtime | ✅ | ✅ | ❌ | ✅ | pcm16/24000 | pcm16/24000 |
| Pipecat | ✅ | ✅ | ❌ | ✅ | pcm16/24000, mulaw/8000, opus | pcm16/24000, mulaw/8000, opus |
| Twilio | ❌ | ❌ | ✅ | ✅ | mulaw/8000 | mulaw/8000 |
| Vapi | ✅ | ✅ | ❌ | ❌ | pcm16/16000 | pcm16/16000 |
| WebRTC | ❌ | ❌ | ❌ | ❌ | pcm16/24000 | pcm16/24000 |
| WebSocket | ❌ | ❌ | ❌ | ❌ | pcm16/24000 | pcm16/24000 |
| Column | Meaning |
|---|---|
streaming_transcripts | Adapter emits incremental transcript events during a turn |
native_vad | Adapter has built-in voice activity detection |
dtmf | Adapter can detect and forward DTMF (keypad) tones |
interruption | Adapter supports barge-in / user-initiated interruption |
input_formats | Audio formats the adapter accepts from the user simulator |
output_formats | Audio formats the adapter sends to the scenario harness |
Wire transport and shipping status
The capabilities table above describes what each adapter supports.
The table below describes how each adapter is wired and whether it is
shipping or still stubbed behind PendingTransportError.
| Adapter | Wire transport | Real I/O? |
|---|---|---|
ComposableVoiceAgent | STT + LLM + TTS pipeline (provider-defined) | ✅ shipping |
ElevenLabsAgentAdapter | WebSocket (ElevenLabs Convai) | ✅ shipping |
GeminiLiveAgentAdapter | WebSocket (Gemini Live) | ✅ shipping |
LiveKitAgentAdapter | WebRTC (LiveKit room) | 🚧 stub (PendingTransportError) |
OpenAIRealtimeAgentAdapter | WebSocket (OpenAI Realtime) | ✅ shipping |
PipecatAgentAdapter | WebSocket (Twilio Media Streams protocol) | ✅ shipping |
TwilioAgentAdapter | Media Streams (WebSocket over Twilio) | ✅ shipping |
VapiAgentAdapter | REST (Vapi outbound) | 🚧 stub (PendingTransportError) |
WebRTCAgentAdapter | WebRTC (datachannel + audio track) | 🚧 stub (PendingTransportError) |
WebSocketAgentAdapter | WebSocket (bring-your-own protocol) | ✅ shipping |
Adapters marked 🚧 raise PendingTransportError on connect() and are tracked
as follow-up issues. Their capability declarations are final (they match the
wire spec); only the transport glue code is pending.
Source of truth
Capability values live in each adapter's capabilities: ClassVar[AdapterCapabilities]
declaration. The canonical source file is
python/scenario/voice/capabilities.py.
The generator script that produces the auto-generated table above is
python/scripts/gen_capability_matrix.py.