单通道语音增强中深度学习方法研究现状与展望

Methods of Deep Learning in Monaural Speech Enhancement:State of Art and Prospects

语音增强是语音信号处理领域一种传统且依然非常活跃的研究分支。单通道语音增强是指从单个麦克风采集的带噪语音中尽可能恢复出干净语音,在移动通信、语音交互、数字助听等领域有重要的应用价值。传统的单通道语音增强技术在处理平稳噪声时已取得较好的增强性能,但在非平稳噪声条件下增强效果依然难以令人满意。近年来,随着人工智能的快速发展,基于深度学习的单通道语音增强在处理非平稳噪声问题方面已取得明显的进展。通过系统梳理单通道语音增强中深度学习方法的发展,并按照技术发展脉络,分基于参数映射、基于生成对抗机制和基于弱监督3个方面进行综述,介绍三类方法的基本原理,分析典型文献的技术思路,总结三类方法的优势与存在的问题,最后对深度学习技术在单通道语音增强领域的发展进行了展望。

Speech enhancement is a traditional and still active research branch in the field of speech signal processing.Monaural speech enhancement refers to the recovery of the clean speech from the noisy speech collected by a single microphone,and has important values in many application fields such as mobile communications,speech interaction and digital hearing aid.The traditional monaural speech enhancement technology has already achieved quite good performance in dealing with stationary noise,but it is still not satisfactory when dealing with non-stationary noise.In recent years,with the rapid development of artificial intelligence,the monaural speech enhancement based on deep learning has made significant progress in dealing with non-stationary noise.In this paper,the monaural speech enhancement based on deep learning is systematically reviewed,with three methods including parameter mapping,generative adversarial mechanism and weak supervision introduced according to the technology development.Moreover,the fundamental principles of the above three methods with the technical schemes from typical references are analyzed.In addition,the advantages and existing problems of the three methods are also summarized.Finally,the prospects for the monaural speech enhancement with deep learning are given briefly.