Speaker Verification
A 1:1 biometric task that evaluates whether two speech samples belong to the same speaker using a similarity score and decision threshold.
Verification vs Identification
Speaker verification tests whether a new sample matches a claimed/reference speaker. It is a 1:1 decision problem, unlike identification which searches across multiple candidate speakers.
Embeddings and Scores
Modern systems can map speech segments to Speaker Embeddings. Two vectors can then be scored with cosine similarity or another backend.
The score is not itself the final decision; a threshold must be selected for the intended data and risk.
Error Trade-Off
A stricter threshold can reduce false accepts while increasing false rejects. Metrics such as FAR, FRR and EER therefore need application context.
Channel, noise, segment duration and domain mismatch can substantially shift performance.