Using cohorts to improve speaker identification

In speaker identification it is the system that declares the identity of the unknown speaker. This is different from speaker verification where the unknown speaker has to first claim an identity and the system will either confirm or deny the claim. There are few practical speaker identification systems due to lower performance, especially when the number of people to be identified increases. In this paper, a speaker identification system is proposed that is based on the speaker verification techniques to improve its performance. It is shown in the paper that using cohort-type system the identification rate can be improved from 65.6% to 81.6% on the full NTIMIT database using 7 utterances to train the system.