基于调制频谱的自适应双通道语音增强算法  

Adaptive two-channel speech enhancement algorithm based on the modulation spectrum

作　　者：朱莉[1] 张戈亮[1] 胡广书[1] 

机构地区：[1]清华大学生物医学工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2013年第5期704-709,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(30970756);清华裕元医学科学研究基金资助项目(20200521)

摘　　要：多通道语音增强能提高语音质量、言语可懂度及言语识别率。然而,现有方法需要假定所需语音入射角度必须为0°。为了克服对入射角的限制,该文提出了一种新的自适应双通道语音增强算法。首先,声源通过由分数阶延时滤波器构成的双通道陷零波束形成器,信号被分成前、后半平面2个部分;再经过调制频谱识别,信号被进一步分成语音与非语音成分;最后,语音成分作为自适应滤波器的输入,非语音成分作为噪声参考输入,经过自适应滤波,从而实现语音增强。实验结果表明:当所需语音的入射角不是0°时,本文算法能不失真地恢复纯音和语音的时频信息,有效抑制噪声干扰。与此同时,该文算法收敛速度快,受步长的影响小。因此,该文算法能更好地符合实际声场的要求,能有效地增强语音。Speech enhancement enhances speech quality,intelligibility,and recognition.However,most current methods assume that the desired speech must have a 0propagation angle.Propagation angles not restricted to 0°can also be analyzed with an adaptive two-channel speech enhancement algorithm developed in this study.The sound sources are separated into front-hemisphere and back-hemisphere components using two-channel null beamforming with fractional delay filters.Then,modulation spectra recognition is used to separate the signals into speech and non-speech components.Finally,the speech is enhanced by an adaptive filter,with the speech component as the input and the non-speech component as the noise reference.Results show that when the angle of the desired signal is not 0°,the algorithm can recover the time-frequency domain characteristics of pure tones and speech without distortion and effectively suppress noise.The algorithm has fast convergence and is robust.Thus,the algorithm is capable of speech enhancement in real acoustic environments.

关 键 词：语音增强 波束形成 调制频谱 自适应滤波 

分 类 号：TN912.16[电子电信—通信与信息系统]