Dereverberation binaural source separation using deep learning

Abstract

This paper reported a deep-learning based binaural separation using gammatone-frequency cepstral coefficient (GFCC) and multi-resolution cochleagram (MRCG) as spectral features and interaural time difference (ITD) and interaural level difference (ILD) as spatial features. A binary mask was estimated by deep neural network (DNN) binary classifier that used the features as a training data and ideal ratio mask (IRM) as a training target. In the experiment, a male speaker as a target speech at azimuth 0o and a female speaker as a masker speech at azimuth 30o, 20o, 10o, and 5o in rooms with 0.32, 0.47, 0.68, and 0.89 s reverberation time (RT60). As a training process, 50 mixtures were used for each condition experiment. The classifier contained two hidden layers, 200 binary neurons and 50 epoch, and Restricted Boltzmann Machine (RBM) was used as pre-training process. The RBM and the classifier learning rate was 1 up to 0.001 from epoch 1 to epoch 50. The sound quality results indicated by 88% intelligibility of STOI method which means that the separated sound is easily understood by the listener. The MOS value is 2.8 which means the sentence is clear but spectral distortion is slightly annoying.