Speaker identification experiments using HMMs

The authors present the results of various speaker identification (SID) experiments performed in a text-dependent mode using speaker-dependent HMMs (hidden Markov models). The main goal of this research was to determine how SID accuracy declines as the number of speakers increases. A set of 963 female speakers were studied. The SID experiments used spoken digit strings for identification. The recognition scores were used as the measure of voice similarity. For the closed set of 963 women, excellent SID accuracy (97.8%) was obtained when only carbon microphone speech was used. When carbon microphone speech was used for training and electret microphone speech was used for testing, the accuracy dropped to 91.1% for a closed set of 122 women and further dropped to 82.7% for a closed set of 308 women.<>