基于一步字典学习的语音增强方法

Speech enhancement method based on one-stage dictionary learning

为了提高增强语音的质量和减少运算时间,提出了一种基于一步字典学习（OS-DL）的去噪方法。首先采用OS-DL算法分别训练出纯净语言和噪声的幅度谱字典,接着采用一致性准则限制的批处理（LARC）算法对含噪语音谱进行稀疏表示,最后用得到的稀疏系数对纯净语音幅度谱进行估计,并结合含噪语音信号的相位重构纯净语音,实现语音增强。对高斯白噪声下不同信噪比的含噪语音进行仿真实验表明,基于OS-DL的语音增强方法在涉及质量的多项客观评价指标上都有较大改进,提升了增强语音的可懂性;与基于K-SVD学习字典的语音增强方法相比,OSDL的语音增强方法消耗的时间减少了150~200 s。

To improve speech quality and reduce computational time, a speech de-noising method based on One Stage Dictionary Leaning （OS-DL） was proposed. Firstly, the magnitude spectrum sub-dictionaries of clean speech and noise were learned form clean speech and noise training databases separately, and the two sub-dictionaries were combined to form a composite dictionary. Then, the noisy speech was sparsely represented over the eomposite dictionary using batch LARS （ Least Angle Regression） with coherence criterion （LARC） algorithm. Finally, the magnitude spectrum of clean speech could be estimated by corresponding sub-dictionary and coding coefficients. The clean speech was reconstructed by the estimated speech magnitude spectrum combining the phase of the noise speech. Simulation experimental results show： in white Gaussian noise conditions, the evaluation indexes of speech are improved, and make the speech better understandable. Meanwhile, compared with K-SVD, the computational time of OS-DL method is reduced by 150 to 200 seconds.