Method based on fractional spectral subtraction for speech enhancement

A method based on fractional spectral subtraction for speech enhancement (FSS) is proposed. It applies FRFT (FRactional Fourier Transform) to noisy speech, the estimated fractional noise spectrum is then subtracted from the derived fractional speech-noise spectrum. Finally, the denoised speech is obtained by inverse fractional Fourier transform. Theory analysis indicates that the new method can find an optimal fractional order, which can best separate speech for noisy data in the fractional Fourier domain. The result is that the performance of the enhanced speech is effectively improved. Computer simulation shows that the SNR improvement amount and Itakura distance decrease amount of the proposed method are superior to those of Spectral Subtraction (SS) for male/female speech corrupted by additive white Gaussian noise. In addition, the music noise of the enhanced speech is remarkably suppressed.