Browse by topic
Type of publication
Comparison of several speaker adaptation alternatives for a Spanish speech task. Actas de las III Jornadas de Tecnologías del Habla, 2004.Comparison of several speaker adaptation alternatives for a Spanish speech task C.D. Martínez-Hinarejos, L. Rodríguez, J.A. Sánchez, A. Sanchis, E. Vidal Nowadays, speaker-independent speech recognition is the most commonly used alternative for building speech recognition systems. With this approach, the parameters of the acoustic models (generally, Hidden Markov Models) can be robustly estimated. However, optimal performance it is not possible for a specific speaker. To avoid this problem, several speaker adaptation techniques have been developed. One of these techniques is Maximum Likelihood Speaker Adaptation (MLLR), which applies adaptation matrices (estimated from adaptation data of the current speaker) to the means of the acoustic models. In this work, we apply MLLR to Spanish acoustic models and compare several ways of applying MLLR. We also apply other adaptation alternatives in order to compare their results with those of MLLR. In general, our experiments demonstrate that applying speaker adaptation techniques provides better results than using general speaker-independent acoustic models.