Crate wavekat_turn 

wavekat-turn

Unified turn detection with multiple backends.

Provides a clean abstraction over turn-detection models that predict whether a user has finished speaking. Two trait families cover the two fundamental input modalities:

• AudioTurnDetector — operates on raw audio frames (e.g. Pipecat Smart Turn)
• TextTurnDetector — operates on ASR transcript text (e.g. LiveKit EOU)

For most use cases, wrap a detector in TurnController to get automatic state tracking and soft-reset logic for VAD integration. See controller for details.

Feature flags

Feature	Backend	Input
pipecat	Pipecat Smart Turn v3 (ONNX, embedded)	Audio (16 kHz)
wavekat-smart-turn	WaveKat language-specialized fine-tunes (ONNX, runtime download)	Audio (16 kHz)
livekit	LiveKit Turn Detector (ONNX)	Text

wavekat-smart-turn implies pipecat and adds an hf-hub runtime dependency. Weights live in wavekat/smart-turn-ONNX and are cached under $HF_HOME/hub/. Set WAVEKAT_TURN_MODEL_DIR to a directory containing <lang>/smart-turn-cpu.onnx to skip the download.

Re-exports

pub use controller::TurnController;

pub use error::TurnError;

Modules

audio

Audio-based turn detection backends.

controller

error

text

Text-based turn detection backends.

Structs

AudioFrame

A frame of audio samples with associated sample rate.

ConversationTurn

A single turn in the conversation, for context-aware text detectors.

StageTiming

Per-stage timing entry.

TurnPrediction

A turn detection prediction with confidence and timing metadata.

Enums

Role

Speaker role in a conversation turn.

TurnState

The predicted turn state.

Traits

AudioTurnDetector

Turn detector that operates on raw audio.

TextTurnDetector

Turn detector that operates on ASR transcript text.