Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

Regression Class Selection and Speaker Adaptation with MLLR in Mandarin Continuous Speech Recognition

Chengrong Li, Jingdong Chen, Bo Xu

National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China

Currently, CDHMM based continuous speech recognition has been widely extended to speaker-independent (SI) system. However, the performance of the SI system is highly dependent on the speakers, especially for Mandarin speech with accent, speaker adaptation becomes crucial important for real application. In this paper, MLLR approach is studied for speaker adaptation in mandarin continuous speech recognition and three approaches for defining regression classes are investigated: the first is based on Chinese phonetic classification, the second is based on statistical information of mixture distribution parameters and the third is based on state duration using segmental information. Other experiments like the effect of adaptation data and mixtures are presented also in the paper. The new variance-based regression class selecting scheme is proposed and has been proved to be effective.


Full Paper (PDF)

Bibliographic reference.  Li, Chengrong / Chen, Jingdong / Xu, Bo (1999): "Regression class selection and speaker adaptation with MLLR in Mandarin continuous speech recognition", In EUROSPEECH'99, 2503-2506.