Spectral Subtraction Based on the Structure of Noise Power Spectral Density

 |  | 
 

It is inevitable to encounter NOISE in daily communication. Noise not only decreases the definition and understandability of the sound, but also brings more burdens to the human ears. In order to reduce noise, speech enhancement algorithm is often adopted to get a clearer sound signal. Spectral subtraction and subspace method are the most important nonparametric method in speech enhancement algorithm.

Owing to its simple theory, stable performance and low arithmetic demand, spectral subtraction is widely used in this research field. However, this method would cause musical noise because of its great estimation variance. Many researchers home and abroad carried out many studies to solve the problem, but still didn't settle it perfectly. So under the guidance of Prof. LI Xiaodong, ZHENG Chengshi, HU Xiaohu, ZHOU Yi of Key Laboratory of Noise and Vibration Research, Institute of Acoustics conducted a series of studies and proposed a novel spectral subtraction algorithm based on structure of noise power spectral density (NPSD-SS), for reducing musical noise without introducing audible speech distortion.

First, they proposes an adaptive averaging periodogram based on the structure of the noise spectrum (NPSD-AAP), where the better performance of the NPSD-SS is achieved mainly due to that the proposed NPSD-AAP provides a low-variance and adaptive-bandwidth spectral estimator. Second, the maximum noise reduction is adaptively determined by the property of the noise spectrum to further suppress the non-continuous noise components. They carry out practical tests and the objective results show that the NPSD-SS is better than the conventional spectral subtraction (CSS) in terms of the signal-to-noise ratio (SNR) improvement and the amount of noise reduction. Informal listening tests further confirm the validity of the proposed NPSD-SS.
Appendix: