基于字典学习和稀疏表示的单通道语音增强算法综述*

如何从带噪语音信号中恢复出干净的语音信号一直都是信号处理领域的热点问题。近年来研究者相继提出了一些基于字典学习和稀疏表示的单通道语音增强算法，这些算法利用语音信号在时频域上的稀疏特性，通过学习训练数据样本的结构特征和规律来构造相应的字典，再对带噪语音信号进行投影以估计出干净语音信号。针对训练样本与测试数据不匹配的情况，有监督类的非负矩阵分解方法与基于统计模型的传统语音增强方法相结合，在增强阶段对语音字典和噪声字典进行更新，从而估计出干净语音信号。本文首先介绍了单通道情况下语音增强的信号模型，然后对4种典型的增强方法进行了阐述，最后对未来可能的研究热点进行了展望。

Review for speech enhancement algorithms based on dictionary learning and sparse representation

How to recover the clean speech signal from the noisy signal has always been a hot issue in the field of signal processing. In recent years, single-channel speech enhancement algorithms based on dictionary learning and sparse representation have been proposed. These algorithms make full use of the sparsity of signals in time-frequency domain and construct the dictionary by learning the structure characteristics of signals. Finally, the clean speech is estimated by projecting the noisy signal in the dictionary. In terms of mismatched training data, a new approach combining the supervised non-negative matrix factorization method with conventional statistical model-based enhancement method has been proposed, which can update the speech and noise dictionaries in the enhancement stage and estimate the clean speech. This paper first introduces the signal model of speech enhancement under single-channel condition, and then expounds four typical enhancement methods. Finally, the future research directions are prospected.