Periodic/aperiodic decomposition for improving coherence based multi-channel speech denoising

In this paper, we are interested in bi-channel speech denoising by exploiting the “similarity” between the observations expressed in term of spectral coherence function. Periodic/aperiodic decomposition, a powerful tool for speech analysis is used to improve speech denoising. More precisely, instead of applying coherence based filter to the couple of observations, we refine noisy observations by decomposing them into two sub-signals. Moreover, we adapt the coherence functions of each couple of sub-signals according to their properties. Experiments carried out on artificially and naturally noisy signals indicate that, relatively to the classical coherence denoising technique, significant gains are achieved by the proposed method.