摘要
面向通信领域广泛使用的线性预测声码器,设计了一种不经过“解码-特征提取”过程,而直接由传输码流截取说话人特征的方法,并针对宽带自适应多码率声码器(AMRWB)建立了与文本无关的话者确认系统.系统采用基于概率统计模型的GMMUBM结构,以LPC倒谱作为主要的话者特征矢量,并加入基音衍生参数以提高确认性能.实验表明,该系统在运算速度提高一个数量级的情况下,达到了与基于重建语音的话者确认系统相接近的性能,且对码率失配具有良好的鲁棒性.
A feature extraction method is designed for linear predict vocoders widely used in the communication field. In this method, feature vectors are extracted not from the decoded waveform, but from the bit stream of transmission directly. Specifically for Wideband Adaptive Multi Rate vocoder （AMR-WB）, we implemented a text-independent speaker verification system. Which employs the probability-statistics-based GMM-UBM framework as speaker model and takes LPC cepstrum and pitch derived parameters as feature vectors. Experiments indicate that the half-decoded based system, which runs ten times faster than the decoded-based system, is capable of similar performance to the latter, and shows robustness for code rate mismatch of AMR-WB.
基金
国家自然科学基金(6027039)
安徽省自然科学基金(01042205)资助项目.
