一种改进窗函数的低时延语音增强算法  

A Low-Latency Speech Enhancement Algorithm Based on Improved Window Function

作　　者：吴君钦[1] 王迎福 WU Jun-qin;WANG Ying-fu(Institute of Information Engineering,Jiangxi University of Science and Technology,Ganzhou Jiangxi 341000,China)

机构地区：[1]江西理工大学信息工程学院,江西赣州341000

出　　处：《计算机仿真》2022年第2期203-211,共9页Computer Simulation

基　　金：国家自然科学基金(61741109)。

摘　　要：语音增强技术作为语音数字信号处理前端的预处理技术,在提高语音的可懂度和总体感知质量等方面扮演着重要角色。而以短时傅里叶变换为基础的语音增强算法会产生窗大小加跳数的算法延时,通常在考虑窗口大小和频谱分辨率后,算法的固有时延会大于64ms。然而,这样的高时延对于包括助听器在内的一些实时性要求较高的应用是很难满足的。为解决上述问题,对传统汉宁窗进行了改进,提出一种非对称的窗函数,并将其与无监督GCC-NMF算法相结合得到了一种无监督的两通道低时延的GCC-NMF语音增强算法。通过从SiSEC获取的语音和现实噪声的两通道混合信号数据集上进行性能评估。使用PEASS和BSS Eval工具包分别使用基于感知、基于SNR进行度量,同时使用STOI和ESTOI对语音的清晰度进行评测。最后,将上述方法与对称窗方法以及其它无监督的语音增强方法进行比较。结果证明,所提出的方法不仅能够将算法时延降低至2ms,还能保持各项评估指标均优于其它无监督以及对称窗方法。As a preprocessing technology for the front end of speech digital signal processing, speech enhancement technology plays an important role in improving speech intelligibility and overall perceived quality. The speech enhancement algorithm based on the short-time Fourier transform can produce an algorithm delay of the window size plus the number of hops. Generally, after considering the window size and the spectral resolution, the inherent delay of the algorithm can be greater than 64 ms. However, such high latency is difficult to meet for some applications that require high real-time performance, including hearing aids.Therefore, based on the traditional Hanning window, this paper proposes an asymmetric window function, and combines it with the unsupervised GCC-NMF algorithm to obtain an unsupervised two-channel low-latency GCC-NMF speech Enhanced algorithm. The performance was evaluated on a two-channel mixed signal data set of speech and real noise obtained from SiSEC. PEASS and BSS Eval toolkits were used to measure based on perception and SNR,while STOI and ESTOI were used to evaluate the intelligibility of speech. Finally, the method was compared with the symmetric window method and other unsupervised speech enhancement methods. The result proves that the proposed method can not only reduce the algorithm delay to 2 ms, but also keep each evaluation index better than other unsupervised and symmetric window methods.

关 键 词：语音增强 非负矩阵分解 非对称窗 广义互相关 低时延 

分 类 号：TN912.35[电子电信—通信与信息系统]