A study on minimum error discriminative training for speaker recognition

The use of discriminative training to construct hidden Markov models of speakers for verification and identification is studied. As opposed to conventional maximum likelihood training which estimates a speaker’s model based only on the training utterances from the same speaker, a discriminative training approach is used which takes into account the models of other competing speakers and formulates the optimization criterion such that speaker separation is enhanced and speaker recognition error rate on the training data is directly minimized. The optimization solution is obtained with a probabilistic descent algorithm. For all experiments an isolated digit database consisting of 100 speakers is used. For speaker identification, the resulting discriminative speaker models reduce the identification error rate by more than 25% over the results obtained with the conventional training algorithm. A new normalized score function is proposed which makes the verification formulation consistent with the minimum error training objective. When combining the proposed verification score function with discriminative training, an average equal error rate of 0.8% is achieved using only one-digit test utterances. This represents an error rate reduction of over 80% from an average equal error rate of 6.1% when using the conventional algorithm for training and the unnormalized score function for testing.