STUDY OF MAXIMUM A POSTERIORI SPEAKER ADAPTATION FOR AUTOMATIC SPEECH RECOGNITION OF PATHOLOGICAL SPEECH
Óscar Saz, Carlos Vaquero, Eduardo Lleida, José Manuel Ortiz-Marcos, César Canalís · 2006
This paper shows the results achieved by the Maxi-mum A Posteriori (MAP) speaker adaptation method in a task of isolated word Automatic Speech Recognition (ASR) system for 6 young speakers with different types of speech pathologies. In this work, the influence of two im-portant variables in speech recognition and speaker adap-tation are studied. On one hand, the performance of the ASR system depending on the acustic unit used by the system (word level unit and context-dependent unit) is analized. On the other hand, how the size of the train-ing set influences the results when using MAP adaptation is studied. The results of this work with the first subset of the Alborada-GTC corpus show that context-dependent units are the most reliable acoustic unit to use, while the use of 2 utterances of the vocabulary recorded in the cor-pus is enough for getting the best results in speaker adap-tation. 1.