基于本征音子说话人子空间的说话人自适应算法

Speaker Adaptation Method Based on Eigenphone Speaker Subspace for Speech Recognition

本征音子说话人自适应算法在自适应数据量充足时可以取得很好的自适应效果,但在自适应数据量不足时会出现严重的过拟合现象。为此该文提出一种基于本征音子说话人子空间的说话人自适应算法来克服这一问题。首先给出基于隐马尔可夫模型-高斯混合模型(HMM-GMM)的语音识别系统中本征音子说话人自适应的基本原理。其次通过引入说话人子空间对不同说话人的本征音子矩阵间的相关性信息进行建模;然后通过估计说话人相关坐标矢量得到一种新的本征音子说话人子空间自适应算法。最后将本征音子说话人子空间自适应算法与传统说话人子空间自适应算法进行了对比。基于微软语料库的汉语连续语音识别实验表明,与本征音子说话人自适应算法相比,该算法在自适应数据量极少时能大幅提升性能,较好地克服过拟合现象。与本征音自适应算法相比,该算法以较小的性能牺牲代价获得了更低的空间复杂度而更具实用性。

The eigenphone speaker adaptation method performs well when the amount of adaptation data is sufficient. However, it suffers from severe over-fitting when insufficient amount of adaptation data is provided. A speaker adaptation method based on eigenphone speaker subspace is proposed to overcome this problem. Firstly, a brief overview of the eigenphone speaker adaptation method is presented in case of Hidden Markov Model-Gaussian Mixture Model （HMM-GMM） based speech recognition system. Secondly, speaker snbspace is introduced to model the inter-speaker correlation information among different speakers＇ eigenphones. Thirdly, a new speaker adaptation method based on eigenphone speaker subspace is derived from estimation of a speaker dependent coordinate vector for each speaker. Finally, a comparison between the new method and traditional speaker subspace based method is discussed in detail. Experimental results on a Mandarin Chinese continuous speech recognition task show that compared with original eigenphone speaker adaptation method, the performance of the eigenphone speaker subspace method can be improved significantly when insufficient amount of adaptation data is provided. Compared with eigenvoice method, eigenphone speaker subspace method can save a great amount of storage space only at the expense of minor performance degradation.