基于i-vector局部加权线性判别分析的说话人识别

I-vector based speaker recognition using local weighted linear discriminant analysis

基于i-vector的说话人识别系统通常采用LDA来消除训练和测试语音之间信道失配,不能保证样本在待识别语音近邻区域内具有最佳的分离度,这就使得目标说话人和其近邻间的得分差异较小,进而导致识别准确性下降。针对该问题,提出基于i-vector局部加权线性判别分析的说话人识别方法(LWLDA)。在计算类内和类间散度时,增加待识别语音近邻样本权重。在此基础上,通过提高待识别语音近邻域局部类间的分辨能力,尽可能减少因信道差异而产生的识别错误。在不同语音库上的实验结果表明:LWLDA在复杂信道环境下能够保持良好的鲁棒性;在交叉信道条件下的识别准确率比LDA平均提高3.6%。

Linear discriminant analysis（ LDA） is often employed to eliminate the channel mismatch between training and testing speeches in identity vector（ i-vector） based speaker recognition systems,which can not provide optimum separation of the samples in the near region of the utterance to be identified. In particular,there is small score difference between the target speaker and corresponding near neighbors,which results in the degradation of recognition accuracy. Aiming at this problem,the i-vector based speaker recognition method with local weighted linear discriminant analysis（ LWLDA） is proposed. In the calculation of inter-class scatter and intra-class scatter,we increase the weights of the samples near the utterance to be identified; based on which,through enhancing the local inter-class discrimination ability in the near region of the utterance to be identified,the recognition errors caused by channel difference are reduced as much as possible. The experiments on different speech databases were conducted. The results demonstrate that,the LWLDA achieves good robustness under complex channel noise environment,and the recognition accuracy ratio is increased by 3. 6% under cross channel conditions compared with that of LDA method.