Automatic Visual Feature Extraction for Mandarin Audio-Visual Speech Recognition

Automatic speech recognition (ASR) by machine has been an attractive research area in past several decades. In recent years, there are many automatic speech-reading systems proposed that utilizing the combination of audio and visual speech features. In this paper, we proposed an automatic visual feature extraction approach to extract the visual features of the lips that can be used in the audio-visual speech recognition system. These features are important to the recognition system, especially in noisy condition. The segmentation of the lip region uses both color and edge information. We then establish a set of visual speech parameters and incorporate them into the recognizer. The WD-KNN classifier is used as the recognition engine in this paper. We present recognition performance using various visual features to explore their impact on the recognition accuracy. These features include the geometric and the motion of the lip. The experimental results based on Mandarin databases demonstrate that the visual information is highly effective for improving the recognition performance.