Ensemble System of Deep Neural Networks for Single-Channel Audio Separation

Speech separation is a well-known problem, especially when there only one sound mixture available. Estimating the Ideal Binary Mask (IBM) solution to this problem. Recent research has focused on supervised classification approach. The challenge of extracting features from sources critical for method. been accomplished by using variety feature extraction models. majority them, however, are concentrated single feature. complementary nature various have not thoroughly investigated. In paper, we propose deep neural network (DNN) ensemble architecture completely explore complimentary diverse obtained raw acoustic features. We examined penultimate discriminative representations instead employing acquired output layer. learned were also fused produce new vector, which was then classified Extreme Learning Machine (ELM). addition, genetic algorithm (GA) created optimize parameters globally. results experiments showed that our proposed system considered and produced high-quality IBM under different conditions.