摘要
从线性预测(LP)残差信号中提出了一种新的特征提取方法,这种特征跟单个的说话人的声道密切相关。不是应用傅立叶技术,而是把HAAR小波变换应用于残差信号,而这种计算更简单。通过把HAAR变换运用于LP残差而获得了一个新的特征HOCOR。为了进一步提高识别性能和训练速度,在识别阶段采用了改进的最大分类错误(MMCE)。实验结果显示采用所提出的新的特征和MMCE取得了较好的识别效果。
A new method to extract discriminative features from the linear prediction （LP） residual signal is proposed, which are closely related to the glottal excitation of individual speaker. Rather than taking Fourier transform, HAAR transform is applied to the residual signal, which is computationally simpler. A novel feature HOCOR is acquired by applying LP residue with HAAR transform. In order to improve the performance of recognition and the training velocity, modified maximum likelihood Error （MMCE） in the recognizing stage is appled. Experiment is showed that a better recognition result is acquired with the proposed novel feature and MMCE.
出处
《科学技术与工程》
2008年第12期3159-3161,3174,共4页
Science Technology and Engineering
