Convolutional neural network and feature transformation for distant speech recognition

Abstract

In many applications, speech recognition must operate in conditions where there are some distances between speakers and the microphones. This is called distant speech recognition (DSR). In this condition, speech recognition must deal with reverberation. Nowadays, deep learning technologies are becoming the the main technologies for speech recognition. Deep Neural Network (DNN) in hybrid with Hidden Markov Model (HMM) is the commonly used architecture. However, this system is still not robust against reverberation. Previous studies use Convolutional Neural Networks (CNN), which is a variation of neural network, to improve the robustness of speech recognition against noise. CNN has the properties of pooling which is used to find local correlation between neighboring dimensions in the features. With this property, CNN could be used as feature learning emphasizing the information on neighboring frames. In this study we use CNN to deal with reverberation. We also propose to use feature transformation techniques: linear discriminat analysis (LDA) and maximum likelihood linear transformation (MLLT), on mel frequency cepstral coefficient (MFCC) before feeding them to CNN. We argue that transforming features could produce more discriminative features for CNN, and hence improve the robustness of speech recognition against reverberation. Our evaluations on Meeting Recorder Digits (MRD) subset of Aurora-5 database confirm that the use of LDA and MLLT transformations improve the robustness of speech recognition. It is better by 20% relative error reduction on compared to a standard DNN based speech recognition using the same number of hidden layers.