Module speaker 

Who is speaking, as a vector: the contract every speaker-embedding engine implements.

An engine turns a window of 16 kHz samples into a unit-norm embedding. Two embeddings of one voice point the same way, so their cosine says whether two stretches of speech came from the same person. What to do with that number is policy, kept out of every engine: Diarizer groups voices and claims enrolled names, VoiceprintTracker judges who holds each turn as a recording is heard, and PyannoteVoices settles who spoke when from the audio itself, afterwards.

Structs

Chunk

Model outputs for one chunk, reusable across clustering threshold sweeps.

Decisionsilero and non-target_family=wasm

What may go out now. Segments still undecided stay held.

Diarizer

An embedding engine with enrolled profiles and anonymous voice clusters.

EnrolledNon-target_family=wasm

A finished enrollment: the profile to store, and the score worth showing the person who just recorded it.

EnrollmentNon-target_family=wasm

Samples gathered so far, and the engine that measures them.

Gatesilero and non-target_family=wasm

Holds decoded segments until the timeline decides them.

Grouping

Clustering thresholds, adjustable without rerunning model inference.

Merged

A voice that turned out to be one already heard.

Profile

One enrolled voice.

ProfileStoreNon-target_family=wasm

The directory of enrolled voices.

PyannoteVoices

Releasedsilero and non-target_family=wasm

A segment let out, and the voice it was heard in.

Revision

What an observation changed about names already written.

SpeakerName

An enrolled person’s name, which is also their filename. Holding one is proof it cannot escape the speakers directory.

Spoken

One stretch of the recording, and whose voice was in it.

Standing

One voice as it stands right now, for a caller watching a session settle.

StoredProfilesNon-target_family=wasm

Timelinesilero and non-target_family=wasm

The turns of one recording so far, and how far into it they reach.

TurnFeedsilero and non-target_family=wasm

Feeds one recording to its speech detector and then to its tracker, on a fork of its own, so turns are cut where the audio pauses, not where a decoder finalizes. The tracker hears only what the detector has scored.

Voiceprint

What one local voice sounded like, embedded through its own frame mask.

VoiceprintTrackersilero and non-target_family=wasm

Cuts one recording into turns and judges each one. The diarizer scores whose voice a window is. The segmenter, when there is one, checks a speaker change inside a turn before the turn splits, which owner-only admission needs.

Enums

Admissionsilero and non-target_family=wasm

Which words a session lets out. The tracker reads it to know what to judge, and the gate reads it to know what to release.

Aheadsilero and non-target_family=wasm

How far TurnFeed::hear_ahead got.

EnrollmentErrorNon-target_family=wasm

Label

An enrolled person, or a voice nobody enrolled, numbered in the order it was first heard.

NoSegmentersilero and non-target_family=wasm

The segmenter of a tracker that never checks a speaker change, so it can name voices but not admit only the owner. Nothing can build one.

OwnerTrackingErrorsilero and non-target_family=wasm

Why no tracker can admit only the owner’s words.

ProfileErrorNon-target_family=wasm

ReplaceNon-target_family=wasm

Whether storing a profile may overwrite one already there.

Role

What an enrolled voice is allowed to be. Only an owner’s voice answers “was that you?”, so a guest can be recognized by name without turning on dictation meant for the person who set the machine up.

Speakersilero and non-target_family=wasm

Whose voice the latest turn is, as judged so far.

SpeakerError

SpeakerNameError

Constants

CLUSTER_THRESHOLD

Similarity threshold for anonymous clusters; short audio windows need a lower gate.

ENROLLED_THRESHOLD

Minimum similarity for claiming a guest’s enrolled name.

MIN_SAMPLESNon-target_family=wasm

Below two samples there is nothing to average and no disagreement to measure.

OWN_VOICE_THRESHOLD

The owner is named at this similarity: below a guest’s bar, so a short turn can carry the name. Whether a turn is the owner’s is read by its length instead, in the speaker tracker.

Traits

Embedder

Turns audio into a comparable voice vector.

Engine

How an engine is constructed from its own configuration. NAME names the engine in stored profiles and errors; Config carries whatever it needs to load.

SpeakerTrackersilero and non-target_family=wasm

Says who spoke when in one recording, reading where anyone spoke from a SpeechMap. A recording’s audio arrives in order, in pieces of any size; the map must have scored a frame before the tracker hears it.

Voices

A running account of who is speaking in a recording.

Functions

attribute_spans

Asks a track who owned each paragraph, skipping the ones nobody timed.

centroidNon-target_family=wasm

Normalizes the elementwise mean of unit-length embeddings. Returns None when the mean has no direction.

consistencyNon-target_family=wasm

Mean pairwise cosine similarity of unit-length embeddings.

normalized

Scales a vector to unit length, so cosines are plain dot products. None when the vector has no direction to keep.

settle

Fills missing labels from neighboring speakers. Utterance openings prefer the following label; other gaps prefer the preceding one. All-None input stays unchanged.

validate_model_fileNon-target_family=wasm

crate::backend::validate_model_file, failing as a SpeakerError.