Speech enhancement based on a novel weighting spectral distortion measure

In this paper, we propose a speech enhancement scheme by only using a priori codebook of speech spectral shapes based on a novel spectral distortion measure. The noise signal is estimated by Minima Controlled Recursive Averaging (MCRA) algorithm instead of using noise spectral shapes stored in trained codebooks. Then, combining with the cross-correlation between the spectra of noisy speech and noise, a weighting spectral distortion measure is proposed for optimizing the spectral shapes and the spectral gains of speech and noise. In order to ensure the smooth of speech spectrum, the line spectrum frequencies (LSF) that describe speech spectral shapes are interpolated linearly. While a priori signal to noise ratio (SNR) information is also contributed to further suppress the noise energy. Finally, the noisy speech is passed through the reconstructed Wiener filter to obtain the enhanced speech. The objective tests show that the proposed method has better performance in restraining the fluctuated background noise than the conventional codebook-based method to a large extent.