论文部分内容阅读
针对以往语音增强算法在非平稳噪声环境下性能急剧下降的问题,基于时频字典学习方法提出了一种新的单通道语音增强算法。首先,提出采用时频字典学习方法对噪声的频谱结构的先验信息进行建模,并将其融入到卷积非负矩阵分解的框架下;然后,在固定噪声时频字典情况下,推导了时变增益和语音时频字典的乘性迭代求解公式;最后,利用该迭代公式更新语音和噪声的时变增益系数以及语音的时频字典,通过语音时频字典和时变增益的卷积运算重构出语音的幅度谱并用二值时频掩蔽方法消除噪声干扰。实验结果表明,在多项语音质量评价指标上,本文算法都取得了更好的结果。在非平稳噪声和低信噪比环境下,相比于多带谱减法和非负稀疏编码去噪算法,本文算法更有效地消除了噪声,增强后的语音具有更好的质量。
Aiming at the problem that the performance of speech enhancement algorithm dropped dramatically under non-stationary noise environment, a new single-channel speech enhancement algorithm is proposed based on time-frequency dictionary learning method. First of all, we propose to use time-frequency dictionary learning method to model the a priori information of the spectral structure of noise and integrate it into the framework of convolution nonnegative matrix factorization. Then, in the case of fixed-frequency time-frequency dictionary, Time-varying gain and speech time-frequency dictionary multiplication iterative solution formula; Finally, the iterative formula is used to update the time-varying gain coefficients of speech and noise as well as the speech time-frequency dictionary. By convolution of the speech time-frequency dictionary and the time- Reconstruct the amplitude spectrum of speech and eliminate the noise interference by using binary time-frequency masking method. The experimental results show that the proposed algorithm achieves better results in evaluating a number of voice quality indicators. Compared with multi-spectral subtraction and non-negative sparse coding denoising algorithms in non-stationary noise and low signal-to-noise ratio environment, the proposed algorithm can eliminate the noise more effectively and the enhanced speech has better quality.