CONTINUOUS SPEECH SEGMENTATION BASED ON A SELF-LEARNING NEURO-FUZZY SYSTEM

For reducing requirement of large memory and minimizing computation co mplexity in a large-vocabulary continuous speech recognition system, s peech segmentation plays an important role in speech recognition syste ms. In this paper, we formulate the speech segmentation as a two-phase problem. Phase 1 (frame labeling) involves labeling frames of speech data. Frames are classified into three types: (1) silence, (2) consona nt and (3) vowel according to two segmentation features. In phase 2 (s yllabic unit segmentation) we apply the concept of transition states t o segment continuous speech data into syllabic units based on the labe led frames. The novel class of hyperrectangular composite neural netwo rks (HRCNNs) is used to cluster frames. The HRCNNs integrate the rule- based approach and neural network paradigms, therefore, this special h ybrid system may neutralize the disadvantages of each alternative. The parameters of the trained HRCNNs are utilized to extract both crisp a nd fuzzy classification rules. In our experiments, a database containi ng continuous reading-rate Mandarin speech recorded from newscast was utilized to illustrate the performance of the proposed speaker indepen dent speech segmentation system. The effectiveness of the proposed seg mentation system is confirmed by the experimental results.