Author : Mandar Diwakar and Dr. R.A. Satao

We propose a relapse approach by means of Deep Neural Network (DNN) for solo discourse partition in a solitary channel setting. We depend on a key presumption that two speakers could be very much isolated in the event that they are not very like one another. A divergence measure between two speakers is then proposed to portray the partition capacity between contending speakers. We exhibit that the separation between speakers of various sexes is sufficiently enormous to warrant a potential detachment. We finally propose a DNN design with double yields, one speaking to the female speaker gathering and the other portraying the male speaker gathering. Prepared and tried on the Speech Separation Challenge corpus our trial results show that the proposed DNN approach accomplishes enormous execution increases over the best in class solo strategies without utilizing specific information about the blended objective and meddling speakers and even outflanks the directed speech based strategy.

Keywords: Deep Neural Network , Speech , Channel Separation