Module segmentation 

Speaker segmentation: powerset-classifier + sliding-window aggregator.

Added in v0.6.

Structs

AggregationConfig

Configuration for aggregation.

Aggregator

Aggregator over sliding-window powerset outputs.

BinarizationConfig

Calibrated binarization of segmentation posteriors, pyannote-style: each speaker’s activity probability (sum of the powerset classes containing the speaker) is thresholded with onset/offset hysteresis, then short active blips are dropped and short gaps bridged.

FrameLabel

Decoded label for a single audio frame.

PowersetDecoder

Stateless decoder; methods are associated functions because no per-instance configuration is needed.

PowersetNative

Hand-written powerset-3.0. Same 10 s / 2 s geometry as the shipping ONNX adapter; inference is SincNet + 4× biLSTM in polyvoice-kernels.

RawSegment

One contiguous segment attributed to a single local speaker index.

WindowOutput

One window’s segmentation output.

Enums

PowersetClass

One of the seven powerset classes, identifying which speakers are active.

SegmentationError

Errors from Segmenter implementations.

Constants

MAX_LOCAL_SPEAKERS

Number of local speakers the powerset scheme addresses: the solo classes {0}, {1}, {2} and the three pair classes built from them.

MIN_AUDIO_SAMPLES

Minimum audio length (16 kHz samples) accepted by Segmenter::segment.

Traits

Segmenter

A speaker segmentation engine — turns raw audio into spans of speech attributed to local speaker indices, with overlap detection.

Functions

binarize_frames

Binarize averaged powerset posteriors into per-frame classes.