基于预处理的DOA估计和基频双输入的语音分割

Speech segmentation based on preprocessing DOA estimation and fundamental frequency dual input

语音分割是语音分离系统的一个重要组成部分,它在信源估计和多说话人环境中的自动语音识别、多声源目标跟踪等许多应用中都起着重要的作用,重叠语音的分割一直都是这项工作的重点。在实际生活中,室内的麦克风采集的语音信号通常都包含混响和噪声信号,它们使接收信号的语音质量变差,影响了波达方向估计特征的精度,导致多声源重叠语音的分割性能下降。针对现有的多声源分割方法对噪声和混响信号鲁棒性差的问题,提出了一种通过预处理来消除语音信号中的明显异常噪声和混响信号的方法。该方法使用广义旁瓣相消器和维纳滤波器实现的后滤波器相结合对原始语音信号进行处理,消除了混响和噪声信号,使语音质量得到了提高,进而使波达方向特征估计更加准确。最后用多假设跟踪同时跟踪说话人的基频特征和波达方向特征来进行分割,以多声源重叠语音为例,对AMI语料库中的16个会议音频进行了统计与分析,结果表明,与未进行预处理的方法相比,平均命中率(HIT)提高了2.10%。

Speech segmentation is an important component of speech separation systems,which plays an important role in many ap-plications such as source estimation and automatic speech recognition in multi-speaker environments,multi-source target tracking,etc.Segmentation of overlapping speech has always been the focus of this work.In real life,the speech signals collected by micro-phones in rooms usually contain reverberation and noise signals,which deteriorate the speech quality of the received signals and af-fect the accuracy of the estimated features of the boda direction,leading to the degradation of the segmentation performance of multi-source overlapping speech.To address the problem that existing multi-source segmentation methods are poorly robust to noise and reverberant signals,a method is proposed to eliminate apparently abnormal noise and reverberant signals in speech signals by pre-processing.The method uses a combination of a generalized parametric phase canceller and a post-filter implemented with a Wiener filter to process the original speech signal,eliminating the reverberant and noisy signals,resulting in improved speech quality and,in turn,more accurate estimation of the direction of arrival features.The segmentation is then performed by tracking the speaker's fun-damental frequency features and direction of arrival features simultaneously using multi-hypothesis tracking.16 conference audios from the AMI corpus are statistically and analytically analyzed with multi-source overlapping speech,and the results show that the average hit rate(HIT)rate is improved by 2.10%compared with the method without pre-processing.