@ottervoice/core

createVoiceSession(config) and createOtterVoiceSession(config) accept the same unified audio-turn configuration. The latter is only a more explicit alias.

`VoiceSessionConfig`

Field	Notes
`mode`	`half_duplex` · `full_duplex` · `push_to_talk` · `streaming_transcript`
`runtime`	Platform adapter
`providers.audioLlm`	Required; a native speech model or server-composed voice stack
`providers.asr`	Optional caption ASR when the audio-turn provider does not transcribe input
`language`	Preferred ASR language
`asrPartial`	Emit `asr_partial` (default `true`)
`turnDetection`	VAD while listening
`interruptionDetection`	Stricter VAD during playback
`policy`	Auto-listen, barge-in, false-interrupt recovery
`audioLlmSystemPrompt` / `audioLlmMaxTokens`	Trusted Node/server sessions only; browser standard mode injects them at the gateway
`audioLlmStartTiming`	`after_audio` for parallel low latency; default `after_asr_final`

Core talks to one audio-turn provider. That provider may be one native speech model or a server-composed ASR → LLM → TTS stack. With transcribesInput: true, Core commits the provider’s input transcript as asr_final; otherwise configure a separate caption ASR.

Full fields: API reference. Source: types.ts.

Low-level LLMProvider and TTSProvider contracts remain available for trusted backends implementing a composite AudioLLMProvider; they are no longer VoiceSessionConfig.providers slots. Browser/app integrations should use server-managed providers and keep business and cost policy at their gateway.

`TurnDetectionConfig`

Field	Default (resolved)	Meaning
`strategy`	`hybrid`	`volume` · `manual` · `hybrid`
`minSpeechMs`	`500`	Voiced time before speech starts
`silenceTimeoutMs`	`1200`	Silence before end-of-turn (try 350–600 for latency)
`maxTurnMs`	`120000`	Hard turn cap
`volumeThreshold`	`0.02`	RMS threshold

`VoiceSessionPolicy`

Field	Meaning
`autoStartListening`	Enter listening after `start`
`allowInterruption`	Allow barge-in
`interruptionTailIgnoreMs`	Ignore speaker tail after a tentative pause (default `200`)
`falseInterruptionSilenceMs` / `TimeoutMs`	False-interruption recovery window
`interruptionCooldownMs`	Cooldown after resuming playback
`postPlaybackVadRearmMs`	Wait for a quiet baseline after playback (default `300`)

Lifecycle

await session.start();
await session.finish();
await session.dispose();

Events: events guide. Error codes and usage live on VoiceSessionEventMap.