Interspeech'2005 - Eurospeech

Lisbon, Portugal
September 4-8, 2005

Rapid Speaker Adaptation for Continuous Speech Recognition Using Merging Eigenvoices

Dong-jin Choi, Yung-Hwan Oh

KAIST, Korea

Speaker adaptation in eigenvoice space is a popular method for rapid speaker adaptation. To improve the performance of the method and to obtain stabilized results, the number of speakerdependent models should be increased and a greater number of eigenvoices should be re-estimated. However, the huge computation time required to find eigenvoices makes these solutions difficult, especially in a continuous speech recognition system. This paper describes a method to reduce computation time by estimating eigenvoices only for supplementary speaker-dependent models and merging them with the used eigenvoices. Experimental results show that the computation time is reduced by 73.7% while the performance is almost the same when the numbers of speaker-dependent models in two sets to be merged are the same.

Full Paper

Bibliographic reference.  Choi, Dong-jin / Oh, Yung-Hwan (2005): "Rapid speaker adaptation for continuous speech recognition using merging eigenvoices", In INTERSPEECH-2005, 297-300.