On a dual-gain approach to noise reduction in the STFT domain

Noise reduction is typically achieved by applying a gain filter to the complex spectrum of the noisy speech signal in the short-time Fourier transform (STFT) domain. However, such an approach does not take into account the noncircularity property of the complex speech spectrum. Recently, a widely linear (WL) filtering framework was developed, which can fully take advantage of the second-order statistics of the noncircular speech spectrum for noise reduction. But the optimal WL filters are more complicated to estimate as the estimation involves the use of interframe information. In this paper, we investigate a dual-gain approach, which achieves noise reduction by applying one gain to filter the real part and another gain to filter the imaginary part of the complex noisy spectrum. We show that this approach can be viewed as a particular case of the WL framework. Compared to the classical method with a single gain, this new approach is shown to be able to achieve better noise reduction performance. Another advantage is that the optimal filters with this approach can be implemented using only the current frame of spectra without the need of the interframe information.