ISCA Archive ICSLP 1994
ISCA Archive ICSLP 1994

Speaker adaptation based on fuzzy vector quantization

Jun'ichi Nakahashi, Eiichi Tsuboka

In this paper, we propose the following four speaker adaptation methods based on Fuzzy Vector Quantization (FVQ), Supervised speaker adaptation based on maximization of likelihood from HMM. Supervised speaker adaptation based on minimization of fuzzy objective function weighted by path probabilities from HMM. Simplified unsupervised speaker adaptation based on minimization of fuzzy objective function. Successive unsupervised speaker adaptation based on minimization of fuzzy objective function.

All these methods adapt the reference codebook using adaptation vector, estimated for a certain criterion. The experiments show that all these methods have better performance in comparison with speaker-independent isolated word recognition, when applied to FVQ/HMM. In particular unsupervised speaker adaptation shows that the word recognition rate for "outlier" speakers has been improved from 84.2 % to 88.9 % using 5 adaptation words.


Cite as: Nakahashi, J., Tsuboka, E. (1994) Speaker adaptation based on fuzzy vector quantization. Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994), 467-470

@inproceedings{nakahashi94_icslp,
  author={Jun'ichi Nakahashi and Eiichi Tsuboka},
  title={{Speaker adaptation based on fuzzy vector quantization}},
  year=1994,
  booktitle={Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994)},
  pages={467--470}
}