A new approach to dereverberation and noise reduction with microphone arrays

In this paper the speech enhancement abilities of a new array-based processor have been tested. The proposed system works in three cascade stages. First, the signals are time aligned with the estimated direction of the desired sound source. Second, the signal is decomposed in its allpass and minimum-phase components using cepstral processing. In this moment, beamforming and liftering in cepstral domain is performed, where the output signal reverberation is reduced. The third part consists in a noise canceller, based in the optimum Wiener filtering and Coherence evaluation. The processor have been tested both with a real database and in simulated conditions of noise and reverberation. To evaluate its performance, we have considered the subjective perception of the output signal and objective measurements of LAR distances, Real Cepstrum distances and Signal to Noise Ratios.