ISCA Archive Interspeech 2006
ISCA Archive Interspeech 2006

A speaker adaptation algorithm using principal curves in noisy environments

Jingying Wang, Zuoying Wang

A new speaker adaptation method of speech recognition is proposed in this paper utilizing principal curves algorithm. The key feature of this method is the construction of a transformation function based on the correlation information between observations of different acoustic states. This is an important a priori information crucial to improving system’s recognition performance. Herein the relationships between the statistics information required choosing the best reconstruction of an audio speech pattern and the codebook state parameters of the new algorithm are described, and then the method is applied to a large database of continuous speech. Experiment results on large vocabulary continuous speech recognition database showed that this new method is superior to MLLR adaptation approach in noisy cases, demonstrating that the principal curves speaker adaptation algorithm successfully exploits the correlation information and improve robustness.

doi: 10.21437/Interspeech.2006-341

Cite as: Wang, J., Wang, Z. (2006) A speaker adaptation algorithm using principal curves in noisy environments. Proc. Interspeech 2006, paper 1374-Tue2BuP.6, doi: 10.21437/Interspeech.2006-341

  author={Jingying Wang and Zuoying Wang},
  title={{A speaker adaptation algorithm using principal curves in noisy environments}},
  booktitle={Proc. Interspeech 2006},
  pages={paper 1374-Tue2BuP.6},