A Robust Visual Feature Extraction based BTSM-LDA for Audio-Visual Speech Recognition

The asynchrony for speech and lip movement is key problem of audio-visual speech recognition (AVSR) system. A multi-stream asynchrony dynamic Bayesian network (MS-ADBN) model is proposed for audio-visual speech recognition. Comparing with multi-stream HMM (MSHMM), MS-ADBN model describes the asynchrony of audio stream and visual stream to the word level. Simultaneously, based on profile of lip implemented by using Bayesian tangent shape model (BTSM), linear discrimination analysis (LDA) is used for visual feature extraction which describes the dynamic feature of lip and removes the redundancy of lip geometrical feature. The experiments results on continuous digit audio-visual database show that lip dynamic feature based on BTSM and LDA is more stable and robust than direct lip geometrical feature. In the noisy environments with signal to noise ratios ranging from 0 dB to 30 dB, comparing with MSHMM, MS-ADBN model with MFCC and LDA visual features has an average improvement of 4.92% in speech recognition rate.