基于高层信息特征的重叠语音检测

发布时间：2018-04-17 00:05

本文选题：重叠语音检测 + 高层信息特征　；参考：《清华大学学报(自然科学版)》2017年01期

【摘要】：重叠语音是影响说话人分割性能的主要因素之一。该文提出了基于语音高层信息特征的重叠语音检测方法以提高说话人分割效果。首先用通用背景模型(universal background model,UBM)提取语音的语言学高层信息特征,并融合这些特征和Mel频率倒谱系数(Mel frequency cepstral coefficient,MFCC)特征建立隐Markov模型(hidden Markov model,HMM)检测重叠语音,然后对处理后的语音进行说话人分割。实验结果表明:对于由TIMIT语音库生成的数据集,该方法对重叠语音检测的错误率比单一采用MFCC特征有显著降低,而且说话人分割性能有明显的提高。
[Abstract]:Overlapping speech is one of the main factors that affect the performance of speaker segmentation.In this paper, an overlapping speech detection method based on high level information features is proposed to improve speaker segmentation.Firstly, the general background model universal background model (UBM) is used to extract the high-level linguistic information features of speech, and these features are fused with Mel frequency cepstrum coefficients (Mel frequency cepstral coefficients) to establish a hidden Markov model, hidden Markov model (HMMM), to detect overlapped speech.Then the speech is segmented.Experimental results show that the error rate of overlapping speech detection based on TIMIT speech corpus is significantly lower than that of single MFCC feature, and the performance of speaker segmentation is improved obviously.
【作者单位】：北京工业大学电子信息与控制工程学院;江苏师范大学物理与电子工程学院;
【基金】：国家自然科学基金资助项目(61471014)
【分类号】：TN912.3

【相似文献】