Speaker adaptation is an important method to increase the recognition rate of speaker independent speech recognizers. For telecommunication applications speaker adaptation must operate unsupervised, without an explicit adaptation phase and must be effective after only a few seconds of input speech. Three algorithms for the adaptation of the parameters of Hidden Markov Models (HMM) are presented. Adaptation of emission probabilities in supervised mode decreases string error rate for a connected-digits-string recognition task by 40 % after two adaptations for each digit. An unsupervised time-synchronous algorithm for adaptation of the emission probabilities decreases errors by 51 %, and unsupervised adaptation of the transition probabilities by 21 %. The simultaneous adaptation both of the emission and the transition probabilities reduces errors by 66 %.
Bibliographic reference. Dobler, S. / Rühl, Hans-Wilhelm (1995): "Speaker adaptation for telephone based speech dialogue systems", In EUROSPEECH-1995, 1139-1143.