字典学习和稀疏表示的无监督语音增强算法  

Unsupervised speech enhancement algorithm based on dictionary learning and sparse representation

作　　者：李轶南[1] 贾冲[1] 张立伟[1] 闵刚[1] 曾理[1] 

机构地区：[1]解放军理工大学指挥信息系统学院,南京210007

出　　处：《计算机应用》2014年第A01期257-261,共5页journal of Computer Applications

基　　金：江苏省自然科学基金资助项目(BK2012510)

摘　　要：针对非结构噪声难以去除的问题,基于字典训练和稀疏表示提出一种无监督语音增强算法。该算法通过构造过完备字典并使用带噪语音样本对其进行训练来实现。首先指出K-奇异值分解算法(K-SVD)存在的不足并提出一种新的改进的字典训练算法:K-双边随机投影算法(K-BRP);然后使用K-BRP算法不断更新字典矩阵和相应的增益系数矩阵,从被非结构化噪声所污染的带噪语音中提取出结构性强的纯净语音。大量实验结果表明,由于训练样本考虑到了语音信号的时频域局部结构特征,所提算法能够很好地消除随机噪声,并且在低信噪比情况下仍然能够保持较高的语音质量和可懂度。To solve the difficulty in enhancing the speech contaminated by unstructured noise, an unsupervised speech enhancement algorithm based on dictionary training and over-completely representation was proposed. In the enhancement stage, the technique alternated between sparse coding of the gain matrix and updating the dictionary atoms by using the K- Bilateral Random Projection （K- BRP） algorithm which is a faster modification of the K- Singular Value Decomposition （K- SVD） algorithm. In this way, clean speech was extracted from the noisy speech. Extensive experimental results show that the proposed algorithm achieves better performance in terms of speech quality and speech intelligibility even in low Signal-to-Noise Ratio （SNR） condition by considering the local time-frequency characteristics of speech.

关 键 词：语音增强 无监督 字典训练 稀疏表示 K-SVD算法 

分 类 号：TN912.3[电子电信—通信与信息系统]