一种基于说话者的无监督语音分割算法

An unsupervised speech segmentation algorithm based on the speaker

手机对话语音中2个说话者之间存在着信道和声学特征上的差异,利用这种差异可以从对话语音中分出属于每个话者的语音部分。文章重点讨论了一种基于距离的无监督语音分割算法,并比较了欧氏距离及广义似然比和持续时间相结合的2种距离测度,后者利用假设检验的似然比来描述2个语音段之间的相似性,通过与文本无关的手机对话语音的话者确认系统实验,表明了它比前者更优越,能较好地检测出绝大部分的说话者改变点,且计算代价也较低。

There are differences between the channels and acoustic characters of the two speakers in cellular conversation,which can be applied to segment the speech of each speaker from the cellular conversation.An unsupervised metric-based speech segmentation algorithm is mainly discussed in this paper.And Euclidean distance measure and the distance measure based on generalized likelihood ratio（GLR） and duration are compared.The latter makes use of the likelihood ratio of hypothesis testing to describe the similarity between two speech segments.The text-independent speaker verification system shows the measure based on GLR and duration is better in verifying segment points with low computation cost.