`ComposableVoiceAgent`

Plug in a STTProvider, an LLM model string, and a TTS model string. Useful when you want to test a hand-rolled voice stack against the same scenario harness as the bundled adapters.

Constructor

python

import scenario
 
agent = scenario.ComposableVoiceAgent(
    stt=scenario.ElevenLabsSTTProvider(api_key=...),
    llm="openai/gpt-5.4-mini",
    tts="elevenlabs/some-voice-id",
    system_prompt="You are a helpful assistant.",
)

stt accepts any STTProvider implementation (voice.OpenAISTTProvider, voice.ElevenLabsSTTProvider, or your own). Note llm differs by SDK: Python takes a model string; TypeScript takes an AI-SDK LanguageModel.

Capabilities

streaming_transcripts	native_vad	dtmf	interruption	input_formats	output_formats
✅	❌	❌	❌	`pcm16/24000`	`pcm16/24000`

VAD: Composable has no native VAD; the SDK falls back to webrtcvad. See Troubleshooting → VAD didn't fire.

Worked examples

See the ElevenLabs composable variant for a working pattern in both SDKs.
TypeScript: composable-stt-swap.test.ts — swaps the judge's STT provider per run via run(cfg, { voice: { stt } }).