Struct PipecatSmartTurn 

pub struct PipecatSmartTurn { /* private fields */ }

Pipecat Smart Turn v3 detector.

Wraps the Smart Turn v3 architecture (Whisper-Tiny encoder + binary classification head). Use new for the embedded upstream weights, or with_variant to pick a WaveKat fine-tune at runtime.

Buffers up to 8 seconds of audio internally. Call push_audio with every incoming 16 kHz frame, then call predict when the VAD fires end-of-speech to get a TurnPrediction.

Usage with VAD

use wavekat_turn::audio::PipecatSmartTurn;
use wavekat_turn::AudioTurnDetector;

let mut detector = PipecatSmartTurn::new().unwrap();
// ... feed frames via push_audio ...
let prediction = detector.predict().unwrap();
println!("{:?} ({:.2})", prediction.state, prediction.confidence);

Implementations

impl PipecatSmartTurn

pub fn new() -> Result<Self, TurnError>

Load the upstream Pipecat Smart Turn v3.2 model embedded at compile time.

Equivalent to with_variant(SmartTurnVariant::PipecatV3).

pub fn with_variant(variant: SmartTurnVariant) -> Result<Self, TurnError>

Load a specific variant of the Smart Turn model.

• SmartTurnVariant::PipecatV3 uses the embedded ONNX bytes — no network required.
• SmartTurnVariant::Wavekat (when the wavekat-smart-turn feature is enabled) downloads the corresponding language file from the wavekat/smart-turn-ONNX HuggingFace repo and caches it under $HF_HOME/hub/. Set WAVEKAT_TURN_MODEL_DIR to point at a pre-populated directory (offline / CI use).

pub fn from_file(path: impl AsRef<Path>) -> Result<Self, TurnError>

Load a model from a custom path on disk.

Useful for CI environments that supply the model file separately, or for evaluating fine-tuned variants without recompiling.

Trait Implementations

impl AudioTurnDetector for PipecatSmartTurn

fn push_audio(&mut self, frame: &AudioFrame<'_>)

Append audio to the internal ring buffer.

Frames with a sample rate other than 16 kHz are silently dropped. The ring buffer holds at most 8 s; older samples are evicted.

fn predict(&mut self) -> Result<TurnPrediction, TurnError>

Run inference on the buffered audio.

Takes a snapshot of the ring buffer, pads/truncates to 8 s, extracts Whisper log-mel features, and runs ONNX inference.

fn reset(&mut self)

Clear the ring buffer. Call at the start of each new speech turn.

impl Send for PipecatSmartTurn

impl Sync for PipecatSmartTurn