A High-Performance Speech-Recognition Method Based on a Nonlinear Neural Network
Phung Hung Binh, Pham Viet Hoang, Dang Xuan Ba · 2021
Speech recognition is an important technology to enable and improve human-human and human-computer interactions. In this paper, we present a high-performance speech recognition method using a nonlinear neural network. Key features of input voices are first extracted and decoded as mel frequency cepstrum coefficients (MFCC). A softmax-ouput-type neural network is next employed to perceive the obscure input information based on the MFCC features. The learning performance of the network is improved by raw and fine adaptation rules using the gradient descent algorithm, in which novel nonlinear leakage functions are adopted. The effectiveness of the designed method is investigated by comparative real-time experiments.