Modulation Domain Spectral Subtraction Combined with Phase Spectrum Compensation

It is well known that the enhanced speech of existing spectral subtraction algorithms has low intelligibility and will introduce "musical noise". In order to resolve these problems, a modulation domain spectral subtraction combined with phase spectrum compensation will be introduced. The noisy speech signal was divided into several subbands by using the first Fourier transform in this algorithm, and the modulation spectrum was got by using the second Fourier transform on every subband at different times. Then, spectral subtraction method was used on the modulation magnitude spectrum of different subbands, the reduction of background noise was significantly improved. And a modulation phase spectrum compensative method was presented by using an antisymmetry function to further suppress the remaining noise and improve speech intelligibility. Both subjective and objective experimental results under some kinds of noise and different signal-to-noise ratios demonstrate that the modulation domain spectral subtraction combined with phase spectrum compensation is more competitive as compared to the existing spectral subtraction algorithms. The method of this paper significantly reduced the background noise and “musical noise” . In addition, the enhanced speech signal keeps favourable intelligibility and improves subjective perception quality.