基于话者分类和HMM的话者自适应语音识别

Speaker Clustering and HMM Based Speaker Adaptation Speech Recognition

本文提出了一种基于话者分类和ＨＭＭ的话者自适应语音识别方法，采用对参考话者聚类、并按话者类分别建立ＨＭＭ模板的策略，对于新注册的用户，系统只需利用其极少量的语音，便可将与之最相近的一类模板指派给新用户，再采用基于谱空间映射的两级自适应方法，使系统自适应到用户的模式下工作．这种方法既提高了识别性能，又降低了自适应的难度，还有利于ＨＭＭ的建立．讨论了话者分类数和自适应语音数据对话者自适应效果及识别性能的影响，提出了一种在自适应语音数据不足情况下仍具有较好自适应效果的基于ＦＶＱ的码本自适应改进算法，该算法还具有对自适应字表不敏感的特点．

in this paper,a speaker adaptation speech recognition based on speaker clustering and HMMs is presented. For a new user,only using some samples of his data and using the two-step speaker adaptation technique based on spectral mapping,the system adapts the nearest original models to the pattern of the user. It can improve recognition performance and reduce adaption difficulty. Also it is easy to build HMM models.The adaptive effect by speaker classification and adaptation training data are discussed.Because the general codebook ad aptation technique is unsatisfactory when the adaptation data is insufficient.A improved FVQ based codebook adaptation algorithm is proposed and good adaptive performance is obtained especially when the adaptive data is insufficient.