An Improved Multi-Band Spectral Subtraction for Enhancing Speech Degraded by Non-Stationary Noises

This paper proposes an improved multi-band spectral subtraction algorithm with the goal of improving the quality of speech signal in various noise environments. In the proposed enhancement algorithm, the whole speech spectrum is divided into different uniformly spaced continuous frequency bands and spectral over-subtraction is performed independently, in each band. The proposed algorithm uses a novel approach to estimate the noise ffom each band continuously, without using speech pause detection. The noise is estimated and updated by adaptively smoothing the noisy signal power in each uniformly spaced frequency band. The smoothing parameter is controlled by a linear function of a-posteriori signal-to-noise ratio (SNR). The experiments are conducted for various types of noises. The results of proposed enhancement algorithm are compared with the reference multi-band spectral subtraction algorithm. To test the performance of the proposed speech enhancement algorithm, objective quality measurement tests (SNR, segmental SNR (Seg.SNR), and perceptual evaluation of speech quality (PESQ)) and spectrogram with informal listening tests are conducted for various noise types at different levels of SNRs. Experimental results and objective quality evaluation test results confirmed the performance of proposed enhancement algorithm. The proposed enhancement algorithm provided sufficient noise reduction and good perceptual quality, without causing considerable signal distortion and remnant musical noise.