Integration of deep learning with expectation maximization for spatial cue-based speech separation in reverberant conditions

In this paper, we formulate a blind source separation (BSS) framework, which allows integrating U-Net based deep learning network with probabilistic spatial machine expectation maximization (EM) algorithm for separating speech in reverberant conditions. Our proposed model uses pre-trained convolutional neural network, U-Net, clustering the interaural level difference (ILD) cues and phase (IPD) cues. The integrated exploits complementary strengths of two approaches to BSS: strong modeling power supervised networks ease unsupervised algorithms, whose few parameters can be estimated on as little single segment an audio mixture. results show average improvement 4.3 dB signal distortion ratio (SDR) 4.3% short time intelligibility (STOI) over EM MESSL-GS (model-based expectation–maximization localization garbage source) 4.5 SDR 8% STOI (U-Net) SONET under conditions ranging from anechoic those mostly encountered real world.