论文部分内容阅读
该文报告了组合LPC参数以及基频F0的高斯混合模型(GMM)电话语音说话人自动识别技术的实验研究结果。该研究在基线试验中GMM使用16混合共分散对角矩阵,特征量为LPC倒谱系数。而在开发系统测试中分别利用语音的全发话区间和有声区间两部分参数增加基频参数进行试验,并给出实验比较结果。在50人电话通话开放集自动切分语音流实验中正确识别率为76.97%,而提案方法为80.29%,改善率为3.32%。接近人工切分语音流时的识别率82.34%。
This paper reports the results of an experimental study of automatic speech recognition technology for Gaussian Mixture Model (GMM) telephone speech with LPC parameters and fundamental frequency F0. In this study, GMM used 16 mixed co-dispersive diagonal matrices in baseline tests, and the feature quantity was LPC cepstral coefficients. In the development system test, we use two parameters of the speech-only interval and the voiced interval respectively to increase the fundamental frequency parameters and give the experimental comparison results. The correct recognition rate of auto-segmentation speech stream experiment in 50 people’s telephone conversation set was 76.97%, while the proposal method was 80.29% and the improvement rate was 3.32%. The recognition rate close to artificially sliced speech stream is 82.34%.