Task typology

 

Speaker identification versus speaker verification

   

As we briefly mentioned in the introduction, speaker recognition   covers two different areas: on the one hand, speaker identification,   on the other hand, speaker verification . As [Doddington (1985)] describes it, the goal of a speaker identification  task is to classify an unlabelled  voice token as belonging to one of a set of n reference speakers, whereas the speaker verification  task is to decide whether or not the unlabelled  voice belongs to a specific reference speaker.

Speaker identification  is therefore a 1 out of n decision, in the case of closed-set speaker identification , the result of which is an identity assignment  to an applicant speaker . However, in practical applications, open-set speaker identification   requires an additional outcome of rejection , corresponding to the possibility that the unlabelled  speech token does not belong to any of the registered speakers . In such circumstances, the applicant speaker   is called an impostor. 

Speaker verification  can be viewed as a particular case of open-set speaker identification,   corresponding to n = 1. The speaker verification  system takes a test voice sample and a claimed identity as input, and returns a binary decision: acceptance  if the applicant speaker  is considered to be the genuine speaker  or rejection  if he is considered to be an impostor  (as regards the claimed identity).

Conversely, open-set speaker identification   can be understood as a step of closed-set speaker identification , followed by a step of speaker verification, the latter using the identity assigned by the former, as the claimed identity.    

Related tasks

Beyond this major distinction between identification and verification, other related tasks can be mentioned.

• Speaker matching , that is, to choose a speaker in a closed-set of references which is most similar to a current speaker, even though it is known in advance that the applicant speaker   is not a registered speaker.  This appears to be a particular case of speaker cluster selection,   where each cluster consists of one speaker only.
• Speaker labelling  , when the identity of speakers taking part in a conversation are registered, and the goal is to localise when their successive interventions begin and end.
• Speaker alignment , when the identity and order of speakers taking part in a conversation are known and the goal is to localise when each of their interventions begins and ends.
• Speaker change detection , when the goal is to detect a change of speaker along a speech stream.

Types of error

For closed-set speaker identification,   a misclassification   error occurs when a registered speaker  is mistaken for another registered speaker (the mistaken speaker).  

For speaker verification , two types of error must be distinguished: false rejection  when a genuine speaker  is rejected and false acceptance  when an impostor is accepted as the speaker he claimed he was (the violated speaker ).

For open-set identification,   the three types of error can occur. Usually, misclassifications  and false acceptances  are considered as equally harmful, and therefore merged together. However, these two types of error may not have the same consequences in some practical applications.