This research report addresses the problem of multichannel audio source separation. We propose a deep neural network (DNN) based framework where the source spectra are estimated using DNNs and used in a multichannel filter. The filter is derived using an iterative expectation-maximization (EM) algorithm, in which spatial covariance matrices encode the spatial information. We present an extensive experimental study to show the impact of different design choices on the performance of the proposed technique. We consider different cost functions for the training of DNNs, namely Itakura-Saito (IS) divergence, Cauchy cost function, phase-sensitive cost function, and mean squared error (MSE). The use of probabilistically motivated cost function, such as the IS divergence, is interesting because it leads to a mathematically rigorous EM interpretation for the proposed framework. We also study the number of EM iterations and the use of multiple DNNs, where each DNN aims to improve the spectra estimated by the preceding EM iteration. Finally, we present its application to a speech enhancement problem. The experimental results show the benefit of the proposed multichannel approach over a single-channel DNN-based approach.

EXPLOR_STEP=$WICRI_ROOT/Wicri/Lorraine/explor/InforLorV4/Data/Main/Exploration

HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 000075 | SxmlIndent | more

HfdSelect -h $EXPLOR_AREA/Data/Main/Exploration/biblio.hfd -nk 000075 | SxmlIndent | more

{{Explor lien
   |wiki=    Wicri/Lorraine
   |area=    InforLorV4
   |flux=    Main
   |étape=   Exploration
   |type=    RBID
   |clé=     Hal:hal-01163369
   |texte=   Multichannel audio source separation with deep neural networks
}}