基于音素后验概率和层次凝聚聚类算法的音素边界检测

摘要

A method of phoneme boundary detection based on phoneme posterior probability and Hierarchical Agglomerative Clustering(HAC) is presented. According to this method, phoneme posterior probabilities should first of all be got by Temporal Pattern(TRAP), and then HAC algorithm is chosen to cluster the phoneme posterior probability. Finally, a reasonable threshold can be obtained by all loss function values, and the number of clusters and the phoneme boundaries can be determined by the threshold. The experimental results show that this method is efficient and bears a good detection performance; and the phoneme posterior probabilities are more suitable for phoneme boundary detection than the Mel-Scale Frequency Cepstral Coefficients(MFCC).%提出了一种基于音素后验概率和层次凝聚聚类算法的音素边界检测方法。该方法首先利用改进的 TRAP结构提取语音信号的帧级音素后验概率；然后，运用层次凝聚聚类算法将提取的音素后验概率进行聚类分析；最后根据其全部的最小损失函数值获取阈值，并通过此阈值决定聚类数目和音素边界。实验证明：该方法具有较好的检测性能，且相对于梅尔倒谱参数(MFCC)，音素后验概率更为适合音素边界的检测。