Browse by topic
Type of publication
Iterative speaker adaptation using MLLR. IV Jornadas en Tecnologías del Habla (IVJTH'2006), 2006.Speech recognition systems are usually speaker-independent, but they are not as good as speaker-dependent systems for specific speakers. An initial speaker-independent system can be adapted to improve recognition accuracy by transforming it into a speaker-dependent system. In this work, a new general acoustic model adaptation technology is presented, using the MLLR algorithm iteratively in a supervised manner. Experiments have been performed on the TT2 Spanish speech corpus. The initial acoustic models were trained from the Albayzin speech database. Their results, which were obtained for 10 speakers, show an improvement in speech recognition accuracy.