Struct PyannoteVoices 

pub struct PyannoteVoices<S, E> { /* private fields */ }

Implementations

impl<S: Segmenter, E: Embedder> PyannoteVoices<S, E>

pub fn new( segmenter: S, embedder: E, enrolled: Vec<Profile>, grouping: Grouping, ) -> Self

pub fn shape(&self) -> &ChunkShape

The chunk width the net answers on.

pub fn hop(&self) -> u64

How far the track advances between chunks. Together with the shape it fixes where every chunk sits, so a caller keeping chunks across runs can tell whether the ones it kept still describe this track.

pub fn hear(&mut self, recording: &RecordingReader) -> Option<Chunk>

The next chunk the net can read, None once the recording holds no more. A chunk that cannot be read is skipped rather than ending the walk: its audio is spent either way.

pub fn group(&mut self, chunk: &Chunk) -> Revision

Places every voiceprint one chunk held, then folds any voices they proved to be one.

pub fn heard_so_far(&self) -> Vec<Standing>

Every voice the track is holding, in the order they were first heard.

A caller sweeping Grouping reads this after each chunk to watch a session settle: how many people the track believes in, and how long it took to stop changing its mind.

impl<S: Segmenter, E: Embedder> PyannoteVoices<S, E>

pub fn retained_from(&self) -> u64

The earliest stream position this track still needs. Its reader may release everything before it, and nothing before it can be re-read.

pub fn skip_to(&mut self, position: u64)

Gives up on the audio before position and starts the next chunk there.

The stretch skipped ends up with no speaker, which is what a track too far behind to read it would produce anyway. Not aligned to the hop: the timeline is indexed off absolute positions, so a new origin costs nothing but the chunks that were never read.

pub fn turns(&self) -> Vec<Spoken>

The timeline this track would draw if nobody asked it about spans: every stretch it heard a voice in, in order. Stretches of two voices at once appear once per voice.

Nothing in the product reads this; the diarization bench scores it.

Trait Implementations

impl<S: Segmenter, E: Embedder> Voices for PyannoteVoices<S, E>

fn attribute(&self, spans: &[Range<u64>]) -> Vec<Option<Label>>

Who owns each span. None where the track cannot say, such as past its reach or over audio nobody could embed.