An Effective Approach for Speech Enhancement by Multi-band MMSE Spectral Subtraction

This paper presents an effective single-channel speech enhancement algorithm that combines the merits of multi-band analysis and minimum-mean-squared-error (MMSE) spectral subtraction. Conventional spectral subtraction algorithm advocates subtraction of the noise spectrum estimate over the entire speech spectrum. Taking that most real-world noise are colored with a nonuniform spectrum into account, we introduce a scale factor which reflects the difference of noise influence in different bands to reduce the color noise. Evaluation results show that our approach can achieve a more significant noise reduction for both white and color noise as compared to existing spectral subtraction speech enhancement algorithm.