Improving Audio-Visual Speech Recognition Using Gabor Recurrent Neural Networks

The performance of speech recognition systems can be significantly improved when visual information is used in conjunction with the audio ones, especially in noisy environments. Prompted by the great achievements of deep learning in solving Audio-Visual Speech Recognition (AVSR) problems, we propose a deep AVSR model based on Long Short-Term Memory Bidirectional Recurrent Neural Network (LSTM-BRNN). The proposed deep AVSR model utilizes the Gabor filters in both the audio and visual front-ends with Early Integration (EI) scheme. This model is termed as BRNNav model. The Gabor features simulate the underlying spatiotemporal processing chain that occurs in the Primary Audio Cortex (PAC) in conjunction with Primary Visual Cortex (PVC). We named it Gabor Audio Features (GAF) and Gabor Visual Features (GVF). The experimental results show that the deep Gabor (LSTM-BRNNs)-based model achieves superior performance when compared to the (GMM-HMM)-based models which utilize the same front-ends. Furthermore, the use of GAF and GVF in both audio and visual front-ends attain significant improvement in the performance compared to the traditional audio and visual features.