基于帧符号化的语音相似性度量方法

摘要

We presented a method to measure similarity of speech by using frame symbolization .Firstly ,remo-ving silence parts from speech segments ,MFCC coefficients were extracted from each frame .Secondly ,MF-CC coefficients were classified by KNN-classification algorithm in terms of k-means clustering results ,and speech signals to do symbolization processing according to the classification .Finally ,speech similarity was computed by using Levenshtein distance .Experiment results show that frame symbolization makes distinction between different speeches are more obvious ,and recognition rate has improved significantly .%提出了将语音帧符号化后度量语音相似性的方法。首先，去除语音段中的静音部分，并提取每帧语音的MFCC参数；其次，将MFCC参数进行 k均值聚类和KNN分类，并根据分类结果对语音信号进行符号化；最后，采用编辑距离计算语音段之间的相似性。实验表明，将语音符号化后，音频之间的可区分性更加明显，识别率也有了明显提高。