语音识别中基于i-vector的说话人归一化研究

摘要

i-vector is an important feature which reflects differences of acoustic characteristics between speakers, and has shown effectiveness in speaker identification and speaker verification. Applies the i-vector method to speaker normalization in speech recognition:extracts the i-vectors of training data and carries out unsupervised clustering using the LBG algorithm. Then performs speaker adaptive training using the cluster information. Speech recognition experiments show that this method can consistantly improve the performance.%i-vector是反映说话人声学差异的一种重要特征，在目前的说话人识别和说话人验证中显示了有效性。将i-vector应用于语音识别中的说话人的声学特征归一化，对训练数据提取i-vector并利用LBG算法进行无监督聚类，然后对各类分别训练最大似然线性变换并使用说话人自适应训练来实现说话人的归一化。将变换后的特征用于训练和识别，实验表明该方法能够提高语音识别的性能。