Speech recognition and models
Speaker recognition
Tying a voice to a known person, not merely telling it apart from other voices.
Speaker recognition is deciding who a voice belongs to. The system compares the audio with a stored voice profile and says: this is Kari.
That is a different thing from diarisation, which only tells the voices apart without knowing who they are. Diarisation gives you “Speaker 1”. Speaker recognition gives you “Kari”.
Two variants
Verification. Is this the person they claim to be? One yes or no. Used for sign-in and for confirming identity in customer service.
Identification. Which of the registered people is this? One answer chosen among many.
Voice profiles are biometric data
This is the practical consequence that is most often overlooked. A stored voice profile used to identify a person uniquely is biometric data under Article 9 of the GDPR, and therefore a special category of personal data with stricter requirements for a lawful basis.
Telling speakers apart in a meeting transcript does not require this: that uses diarisation, which stores no profile and identifies nobody. The distinction is worth knowing before you choose a solution, and before you answer a question from your data protection officer.