Deep Neural Network Based Speech Recognition Systems Under Noise Perturbations

Automatic speech recognition, which plays an important role in human-computer interactions, is the cornerstone of communication between human and smart devices. In the past few years, deep neural networks (DNNs) have been deployed in automatic speech recognition with great success. However, recent research has discovered that DNNs are not robust against small perturbations. In this work, we investigate the capability of noise immunity in various neural network models through the speech recognition task. When the noise is introduced into the original speech audio, our experimental results demonstrate that the phoneme error rate (PER) degrades as the signal-to-noise ratio (SNR) reduces across all evaluated neural network models. On the other hand, when the noise is introduced into the Mel-frequency cepstral coefficient (MFCC) features, the multilayer perceptron (MLP) network model outperforms all other recurrent neural network (RNN) models.