A taxonomy of speaker recognition systems

 

 

Any speaker classification or recognition system functions in 3 modes:

• A training mode , during which speaker or speaker class models are built, estimated and stored in the system, as reference patterns. Alternative terms: learning mode, registration, enrolment, subscription, etc.
• A test mode , during which the system performs the recognition of an utterance to be identified (or verified). Alternative terms: recognition mode, trial mode, operating mode, etc.
• An untraining mode , during which a speaker or speaker class model is removed from the list of reference patterns. Alternative terms: unlearning mode, unregistration, subscription cancellation, etc.

Naturally, a given speaker cannot be identified or verified in the test mode  before he has been registered during a training phase . However, training  is usually incremental in several ways. Firstly, a speaker recognition  system is usually operational before all possible speakers are referenced, and new users can be added while the system is already functioning in test mode  for those who are registered. Secondly, some system implementations use speech material obtained during a test phase  to update training references (when the recognition decision is judged reliable enough).

Unlike training  and testing , untraining  does usually not require active participation of the speaker being unregistered.

 
• Task typology
• Levels of text dependence
• Interaction mode with the user
• Definitions
• Examples