Bi-gram constraint segment-based speech modeling : An application to the Japanese mono-syllables recognition

概要

論文の詳細を見る
A new type of stochastic modeling instead of the HMM for speech recognition, named as bi-gram constraint segment-based speech modeling (BSSM) that uses the correlation between two successive VQ codewords is presented. The main trait of this proposed approach is that the model dose not employ the concept of states as in HMM, but that of segments, corresponding to the quasi-stationary elements in speech. Since the segmentation can easily be performed apart from the function of both the training and the testing processors, the model parameters of BSSM can be estimated without any iterative task which has to be unavoidably followed in the case of HMM-based approach. Through experiments for the speaker independent Japanese mono-syllables recognition, our approach shows its outperformance over the HMM-based approaches viz. bi-gram constraint HMM and temporal correlative HMM. Moreover we have succeeded to reduce tremendously the complexities on both computation and storage compared with the HMM-based approaches.
社団法人映像情報メディア学会の論文
1997-09-19