音素层特征超矢量的说话人识别性能及优化

Study on performance and improvement of speaker recognition of phoneme feature supper vector

音素层特征等高层信息的参数由于完全不受信道的影响,被认为可对基于声学参数的低层信息系统进行有益的补充,但高层信息存在数据稀少的缺点。建立了基于音素特征超矢量的识别方法,并采用BUT的音素层语音识别器对其识别性能进行分析,进而尝试通过数据裁剪和KPCA映射的方法来提升该识别方法的性能。结果表明,采用裁剪并不能有效提升其识别性能,但融合KPCA映射的识别算法的性能得到了显著提升。进一步与主流的GMM-UBM系统融合后,相对于GMM-UBM系统,EER从8.4%降至6.7%。

As being hard to be influenced by the channel situation,the higher level information,such as phoneme feature,is recognized to be a good complementarity to the current speaker recognition technology based on lower level information,such as acoustic information.However,the higher level speech information has their inherent limitations of data sparsity.Based on the BUT speaker recognizer platform,the performance of the speaker recognition method based on phoneme feature super vector is analyzed and evaluated.The method of data pruning and Kernel Principal Component Analysis（KPCA） are introduced to improve its recognition performance.Results show that the recognition performance is not effectively improved by the data pruning method,but is greatly enhanced when the KPCA is used.Furthermore,when the current system is integrated with GMM-UBM（Gaussian Mixture Model-Universal Background Model） system,the EER（Equal Error Rate） of the GMM-UBM system can be lowered down from 8.4% to 6.7%.