Speech Segmentation and Clustering Problem Based on Fuzzy Rules and Transition States

摘要

Most of the speech segmentation works are based on the thresholds of parameters to segment the speech data into phonemic units or syllabic units. In this paper, we formulate the thresholds decision as a clustering problem. Feature parameters extracted from the analysis frame is clustered into three types: silence, consonants, and vowels, thereby labeling the speech data. Distributed fuzzy rules which have been used in clustering the numerical data is used for this task. The distributed fuzzy rules, which needn't much training data, have good performance in clustering problem, is thus beneficial for clustering the features of speech data. Four speakers' continuous reading-rate Mandarin speech are given to illustrate this approach to label the speech data. For segmenting the continuous speech data into syllabic units, the segmentation is based on the labeling sequence of the speech data and the transition states of structures in Mandarin words. Effectiveness of this approach has been substantiated by segmentation experiments for continuous, radio news speech samples uttered by two female and two male.