Type alias AgentSessionOptions<UserData>

AgentSessionOptions<UserData>: {
    aecWarmupDuration?: number | null;
    connOptions?: SessionConnectOptions;
    expressive?: boolean | ExpressiveOptions;
    forwardAudioIdleTimeout?: number;
    keytermsOptions?: KeytermsOptions;
    llm?: llm.LLM | llm.RealtimeModel | LLMModels;
    maxToolSteps?: number;
    preemptiveGeneration?: boolean;
    stt?: stt.STT | STTModelString;
    toolHandling?: ToolHandlingOptions;
    tools?: ToolContextLike<UserData>;
    transcriptionTimeout?: number | null;
    tts?: tts.TTS | TTSModelString;
    ttsReadIdleTimeout?: number;
    ttsTextTransforms?: readonly TextTransform[] | null;
    turnDetection?: TurnDetectionMode;
    turnHandling?: Partial<TurnHandlingOptions>;
    useTtsAlignedTranscript?: boolean;
    userAwayTimeout?: number | null;
    userData?: UserData;
    vad?: VAD | null;
    voiceOptions?: Partial<VoiceOptions>;
}

Type Parameters

  • UserData = UnknownUserData

Type declaration

  • Optional aecWarmupDuration?: number | null

    Duration in milliseconds for AEC (Acoustic Echo Cancellation) warmup, during which interruptions from audio activity are suppressed. Set to null to disable. Defaults to 3000, or null for outbound SIP calls.

  • Optional connOptions?: SessionConnectOptions
  • Optional expressive?: boolean | ExpressiveOptions

    Let the LLM steer how the agent sounds.

    When enabled, the provider's markup guide is injected into the LLM prompt so it can emit inline delivery tags (emotion, pacing, non-verbal sounds), which are rendered by the TTS and stripped from the transcript. Pass an ExpressiveOptions object to steer or override the injected instructions. Requires an TTS with a model that declares a markup dialect; it stays off otherwise.

    Default Value

    false
  • Optional forwardAudioIdleTimeout?: number

    Maximum time in milliseconds to wait for the next frame while forwarding TTS audio to the audio output inside performAudioForwarding. Applies to every read, including the first. If exceeded, forwarding is forcibly closed and a stall warning is logged.

    Default Value

    10000
  • Optional keytermsOptions?: KeytermsOptions

    Keyterm biasing for the STT. Holds static keyterms plus keytermDetection (LLM extraction). Applies to STTs that accept a term list; on others it warns and is ignored.

  • Optional llm?: llm.LLM | llm.RealtimeModel | LLMModels
  • Optional maxToolSteps?: number
  • Optional preemptiveGeneration?: boolean

    Deprecated

    Use turnHandling.preemptiveGeneration instead. When set, migrated into turnHandling.preemptiveGeneration.enabled.

  • Optional stt?: stt.STT | STTModelString
  • Optional toolHandling?: ToolHandlingOptions
  • Optional tools?: ToolContextLike<UserData>
  • Optional transcriptionTimeout?: number | null

    Emit user_transcription_timeout when VAD detects user speech during the user's turn but no non-empty final transcript arrives within this many milliseconds after the speech ends. This can happen because STT failed or because audio was intentionally withheld from STT, such as during AEC warmup or uninterruptible agent speech. A non-empty final transcript satisfies the timeout for the current turn even if adaptive interruption detection later discards it as a backchannel. Requires both VAD and STT. Set to null to disable.

    Default Value

    null
  • Optional tts?: tts.TTS | TTSModelString
  • Optional ttsReadIdleTimeout?: number

    Maximum time in milliseconds to wait for the next frame on the TTS audio stream inside performTTSInference. Applies to every read, including the first. If exceeded, the TTS stream is forcibly closed and a stall warning is logged.

    Default Value

    10000
  • Optional ttsTextTransforms?: readonly TextTransform[] | null

    Transforms to apply to TTS input text. Built-in transforms are filter_markdown and filter_emoji; pass null to disable text transforms.

  • Optional turnDetection?: TurnDetectionMode

    Deprecated

    use turnHandling.turnDetection instead

  • Optional turnHandling?: Partial<TurnHandlingOptions>

    Configuration for turn handling.

  • Optional useTtsAlignedTranscript?: boolean
  • Optional userAwayTimeout?: number | null

    If set, set the user state as "away" after this amount of time after user and agent are silent. Set to null to disable.

    Default Value

    15.0
  • Optional userData?: UserData
  • Optional vad?: VAD | null

    Voice Activity Detection. When omitted, AgentSession auto-provisions a bundled inference.VAD({ model: 'silero' }) and marks it as the default (so sites that check whether the user supplied a VAD treat the bundled one as absent). Pass null to opt out entirely.

  • Optional voiceOptions?: Partial<VoiceOptions>

    Deprecated

    use top-level SessionOptions fields instead

Settings

Member Visibility

Theme