Two-stage data-driven single channel speech enhancement with cepstral analysis pre-processing

Most single channel speech enhancement approaches introduce distortion in the speech and create musical noise. In this paper, we propose a novel data-driven two-stage speech enhancement approach based on cepstral smoothing method. By utilizing the cepstral smoothing method and modified recursive a priori and a posteriori signal-to-noise ratio estimation, the main speech information is preserved while noise is attenuated at the first stage. At the second stage, the residual noise is suppressed. The lookup table technique is also applied to achieve better noise suppression. A notable improvement of the proposed approach over the Voice-Activity-Detector (VAD) based decision directed approach is that it introduces less speech distortion while suppressing the noise. Quantifiable results shows that up to 20 percent improvement in terms of speech quality is achieved under certain noisy conditions.