一种基于CASA的单通道语音增强方法

A Novel Single- Channel Speech Enhancement Method Based on Computational Auditory Scene Analysis

基于对计算听觉场景分析(Computational Auditory Scene Analysis,CASA)算法思想的研究,提出了一种单通道语音增强方法。通过分析白噪声、风噪声、周期性噪声三类典型噪声和一般语音信号的频谱特点,构造适合的信号提取特征作为线索,判别出信号时频单元中的主要信号成分,然后对各时频单元乘以相应的衰减系数以掩蔽噪声成分。对仿真实验结果的客观测试和非正式听音测试表明,相对于常用的多子带谱减法和维纳滤波法,所提出的算法能够更有效地抑制白噪声、风噪声、周期性噪声等背景噪声。

A novel single - channel speech enhancement method based on Computational Auditory Scene Analysis is pres- ented. First of all, the time - frequency characteristics of the speech signal and three typical noise including white noise, wind noise and periodic noise are analyzed. Appropriate features as the cues for judging the main component of each time-frequen- cy unit are extracted. After this each time-frequency unit will be multiplied by a masking coefficient to mask the noise com- ponent and enhance the speech component. Objective measurements combined with informal subjective listening tests show that, compared with the popular algorithms such as multiband spectral subtraction and Wiener filtering, the proposed method can suppress background noise such as white noise, wind noise and periodic noise more effectively.