An incremental speaker-adaptation technique for hybrid HMM-MLP recognizer

One of the problems of speaker-independent continuous speech recognition systems is their inability to cope with the inter-speaker variability. When we find test speakers with different characteristics from the ones presented in the training pool we observe a large degradation on the system performance. To overcome this problem speaker-adaptation techniques may be used to provide near speaker-dependent accuracy. In this work we present a speaker-adaptation technique applied to a hybrid HMM-MLP system for large vocabulary, continuous speech recognition. This technique is based on an architecture that employs a trainable linear input network (LIN) to map the speaker specific features input vectors to the speaker-independent system. This speaker-adaptation technique is evaluated in an incremental speaker-adaptation task using a Wall Street Journal (WSJ) database. Both supervised and unsupervised modes are evaluated. The results show that speaker-adaptation within the hybrid framework can substantially improve system performance.