Glossary

embedding

A fixed-dimensional vector representing a speech segment or speaker. Embeddings are commonly compared to produce verification scores.

enrollment

The speech data or embedding(s) used to create a model representation for a claimed identity.

model

In verification tables, the identifier for an enrollment representation. A model can be built from one or more enrollment segments.

segment

A labeled portion of a recording, usually identified by a unique segment id. A segment may be a complete utterance or a time interval in a larger recording.

recording

A source audio, image, or video asset from which one or more segments can be drawn.

trial

One comparison between an enrollment model and a test segment.

trial index

A hyperion.utils.TrialNdx specifying which model/segment pairs should be scored.

trial key

A hyperion.utils.TrialKey that labels valid trials as target, non-target, and optionally spoof.

target trial

A trial in which the test segment belongs to the claimed enrollment identity.

non-target trial

A trial in which the test segment does not belong to the claimed enrollment identity.

score

A numeric measure of similarity or log-likelihood ratio for one trial. Higher values usually indicate stronger evidence for the target hypothesis, subject to the scoring backend’s documented convention.

cohort

A set of background embeddings used for score normalization methods such as adaptive S-norm.

VAD

Voice activity detection: a mask, table, or set of time intervals that indicates speech-active regions.

ark/scp

Kaldi-compatible storage conventions. An ark is an archive of keyed data; an scp is an index that maps keys to archive locations.

EER

Equal error rate: the operating point at which false-reject and false-accept rates are equal. Lower is better.