Struct PowersetSegmenter 

pub struct PowersetSegmenter { /* private fields */ }

ONNX-backed powerset speaker segmenter.

Implementations

impl PowersetSegmenter

pub fn new(model_path: impl AsRef<Path>) -> Result<Self, SegmentationError>

{ true } pub fn new(model_path: impl AsRef<Path>) -> Result<Self, SegmentationError> { true } Load the ONNX model from model_path with the target’s default execution provider (today’s behavior: CoreML on Apple Silicon).

pub fn with_config( model_path: impl AsRef<Path>, config: PowersetConfig, ep: ExecutionProvider, ) -> Result<Self, SegmentationError>

{ true } pub fn with_config( model_path: impl AsRef<Path>, config: PowersetConfig, ep: ExecutionProvider, ) -> Result<Self, SegmentationError> { true } Load with explicit configuration and execution provider.

When the active inference backend is tract (POLYVOICE_INFERENCE_BACKEND=tract or crate::onnx::InferenceBackend::force), this remaps a shipping powerset path to powerset_fp32_tract.onnx if present beside it, and forces micro-batch N to 1 (tract LSTM Scan cannot run N>1). Session pool stays configurable — that is the tract parallelism knob.

pub fn config(&self) -> &PowersetConfig

{ true } pub fn config(&self) -> &PowersetConfig { ret == &self.config }

pub fn model_path(&self) -> &Path

{ true } pub fn model_path(&self) -> &Path { ret == self.model_path.as_path() }

Trait Implementations

impl Segmenter for PowersetSegmenter

fn segment(&self, audio: &[f32]) -> Result<Vec<RawSegment>, SegmentationError>

Segment audio. Audio must be 16 kHz mono f32 PCM.

fn max_local_speakers(&self) -> usize

Max number of distinct local speakers this implementation can output. powerset-3.0 ⇒ 3.

fn supports_overlap(&self) -> bool

True if the implementation can detect overlap (two simultaneous speakers). powerset-3.0 ⇒ true.