On comparison of deep learning architectures for distant speech recognition

Abstract

Deep Learning technologies are becoming the major approaches for natural signal and information processings, including for speech recognition. Many architectures for deep learning have been proposed for automatic speech recognition (ASR). In this paper, we investigate the robustness of various deep learning architectures: DBN-DNN (Deep Belief Network Deep Neural Network), LSTM (Long Short Term Memory), TDNN (Time Delay Neural Network), and CNN (Convolutional Neural Network), for distant speech recognition. The architectures are evaluated on Meeting Recorder Digits (MRD) set of Aurora-5 dataset, a corpus of real recordings on reverberant conditions. Experimental results show that CNN consistently offer the best performance on two most commonly used features in ASR, i.e. Mel Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP).