Optimization of Deep Neural Network for Automatic Speech Recognition

Neural network has achieved improvements in various fields of image processing, natural language processing, object recognition, and acoustic signal classification in automatic speech recognition system (ASR). ASR is the task of mapping the speech signals into the corresponding text without the involvement of human. Recently, paradigm has been shifted from GMM-HMM to Deep Neural Network for speech recognition process. Performance of ASR system depends on how accurately it recognizes the acoustic signals. The recognition rate is directly related to training process of the Deep Neural Network (DNN) i.e. how accurately weights are adjusted in matrix. It short, it can be said that more fine training more accurate results. Therefore, there is a need to propose such technique that optimizes the weight matrix of neural network. In this paper, Meta-heuristic algorithm pigeon inspired optimization (PIO) technique is proposed to optimize weight matrix of DNN model. This technique optimizes the weight matrix using heuristic available. By this way, training time of DNN reduces and recognition rate of system also increases. The result of optimization of weight matrix is evaluated on TIMIT database for phoneme recognition.