论文部分内容阅读
文章提出一种基于小字典训练和过完备稀疏表示的语音增强算法。该算法通过构造过完备的小字典并使用带噪语音的幅度谱对其进行训练来实现。训练过程中通过不断地使用K-SVD算法更新字典矩阵和相应的稀疏系数矩阵来实现对纯净语音的提取,达到语音增强的效果。该方法不同于传统增强算法需要对噪声进行估计与抑制,而是通过稀疏表示将纯净语音从带噪语音中分离出来。主客观测试结果表明,本文方法较好地消除了随机噪声,低信噪比情况下增强效果明显优于传统算法,且能够避免产生音乐噪声。
This paper proposes a speech enhancement algorithm based on small dictionary training and overcomplete sparse representation. The algorithm is implemented by constructing a well-formed dictionaries and training them using the amplitude spectrum of noisy speech. In the training process, the K-SVD algorithm is used to update the dictionary matrix and the corresponding sparse coefficient matrix to extract the pure speech and achieve the effect of speech enhancement. This method is different from traditional enhancement algorithms which need to estimate and suppress noise. Instead, sparse representation separates pure speech from noisy speech. The results of subjective and objective tests show that the proposed method can eliminate random noise well, and the enhancement effect is obviously better than the traditional algorithm under low signal-to-noise ratio and can avoid the generation of music noise.