一种基于非参数贝叶斯理论的语音增强算法

发布时间：2018-06-12 05:07

本文选题：稀疏表示 + 非参数贝叶斯　；参考：《厦门大学学报(自然科学版)》2017年03期

【摘要】：提出一种基于非参数贝叶斯理论的语音增强算法,在稀疏表示的框架下,把字典学习、稀疏系数表示和噪声方差估计融合成一个贝叶斯后验估计的过程,并利用Spike-Slab先验加强稀疏性.首先,将带噪语音分解为干净语音、高斯噪声和残余噪声3个子信号,分别对该3种子信号采用不同的先验概率模型表达,接着采用马尔科夫链-蒙特卡洛算法计算出3个模型中每个参数对应的后验概率,最后基于稀疏表示的框架重构出干净语音.实验数据使用NOIZEUS语音库,采用PESQ和SegSNR作为质量评价指标,分别在信噪比为0,5和10dB的高斯白噪声、火车噪声和街道噪声上验证了其可行性,并与多种常用语音增强方法进行对比,发现其在低信噪比非平稳噪声情况下的增强效果更为理想.
[Abstract]:A speech enhancement algorithm based on non-parametric Bayesian theory is proposed. Under the framework of sparse representation, dictionary learning, sparse coefficient representation and noise variance estimation are fused into a Bayesian posteriori estimation process. Spike-Slab is used to enhance sparsity. Firstly, the noisy speech is decomposed into three sub-signals: clean speech, Gao Si noise and residual noise, and the three subsignals are expressed with different priori probability models. Then the Markov chain-Monte Carlo algorithm is used to calculate the posteriori probability of each parameter in the three models, and the clean speech is reconstructed based on the sparse representation framework. The experimental data are used in Noizus language base, PESQ and SegSNR are used as the quality evaluation indexes. The feasibility of the proposed method is verified on the Gao Si white noise, train noise and street noise with a signal-to-noise ratio of 0 ~ 5 dB and 10 dB, respectively, and is compared with a variety of common speech enhancement methods. It is found that the enhancement effect is better under the condition of low signal-to-noise ratio (SNR) non-stationary noise.
【作者单位】：厦门大学信息科学与技术学院;华南理工大学数学学院;
【分类号】：TN912.35

【相似文献】