融合数据预滤波和频谱展宽的骨导语音增强方法  

Bone-conducted Speech Enhancement by Combining Data Prefiltering with Spectrum Extension

作　　者：张玥 邦锦阳 孙蒙 张雄伟 ZHANG Yue;BANG Jinyang;SUN Meng;ZHANG Xiongwei(College of Command&Control Engineering,Army Engineering University of PLA,Nanjing 210007,China)

机构地区：[1]陆军工程大学指挥控制工程学院,江苏南京210007

出　　处：《陆军工程大学学报》2022年第4期21-29,共9页Journal of Army Engineering University of PLA

基　　金：国家自然科学基金(62071484)。

摘　　要：骨导语音具有天然的抗环境噪声能力,然而,受骨导麦克风佩戴位置和方式的影响,骨导语音在采集过程中常混入骨导麦克风与皮肤或衣服之间的摩擦声,导致现有基于深度学习的骨导语音增强方法鲁棒性不高、适应性不强。为提高骨导语音增强的鲁棒性,提出一种融合数据预滤波和频谱展宽的骨导语音增强方法。该方法首先通过低通滤波对骨导语音数据进行预处理以去除高频噪声,然后对预滤波后的骨导语音进行时频变换,并分别基于U-Net和CRNN两种深度网络进行频谱展宽,最后通过时频逆变换重构出全频带语音。仿真结果表明,与现有深度网络增强方法相比,所提方法可以取得更好的PESQ和STOI客观评价指标,主观听感具有更好的清晰度,且对不同说话人具有更好的适应性。Bone-conducted(BC)speech has the natural immunity to environmental noises.However,due to the influence of the positions and ways of wearing the BC microphone,the friction noise produced between the BC microphone and the skin or clothes is often mixed in the process of BC speech acquisition.As a result,the existing BC speech enhancement solutions are not robust and adaptable enough.In order to improve the robustness of BC speech enhancement,this paper proposed a robust enhancement method of BC speech based on deep learning by combining data prefiltering with spectrum extension.With this method,the BC speech was preprocessed at first by low-pass filtering in order to reduce high frequency noise,and then time-frequency transform was performed on the prefiltered BC speech,and spectrum extension was carried out based on U-Net and CRNN respectively.Finally,full-band speech was reconstructed by inverse time-frequency transform.Simulation results show that compared with the existing deep network enhancement methods,the method proposed in this paper can achieve higher objective evaluation scores of PESQ and STOI and better speech intelligibility,and can also improve the adaptability to different speakers.

关 键 词：骨导语音增强 数据预滤波 频谱展宽 深度学习 

分 类 号：TN912.3[电子电信—通信与信息系统]