论文信息 - Syllable recognition using syllable-segment statistics and syllable-based HMM

Syllable recognition using syllable-segment statistics and syllable-based HMM

In our previous research, we demonstrated the validity of segmental unit input hidden Markov model (HMM), which regards successive four frame MEL-cepstrum coefficients as a feature vector. The vector is reduced to lower dimensions using the KL transform. However, the model considers only the correlation between frames in a short section, but not the correlation between the frames over a long section. In this paper, in order to represent the correlation over a long distance, we use the syllable-segment statistics that are calculated by the concatenation of feature vectors, corresponding to each state in a syllable based HMM. By combining this approach with a segmentalunit input HMM, the syllable recognition rate was improved to 87% from 83% for syllables taken from continuous speech, without using a language model. We also showed the effectiveness for continuous speech recognition.

Seiichi Nakagawa | Nobutoshi Takahashi

[1] Nobuaki Minematsu,et al. Continuous speech recognition using segmental unit input HMMs with a mixture of probability density functions and context dependency , 1998, ICSLP.

[2] Xiaodong Sun,et al. Speech recognition using hidden Markov models with polynomial regression functions as nonstationary states , 1994, IEEE Trans. Speech Audio Process..

[3] Seiichi Nakagawa,et al. Comparative evaluation of segmental unit input HMM and conditional density HMM , 1995, EUROSPEECH.

[4] Yung-Hwan Oh,et al. A segmental-feature HMM for speech pattern modeling , 2000, IEEE Signal Processing Letters.

[5] Seiichi Nakagawa,et al. Evaluation of segmental unit input HMM , 1996, 1996 IEEE International Conference on Acoustics, Speech, and Signal Processing Conference Proceedings.