Study of different feature extraction method for visual speech recognition

Abstract

The prime rule of visual speech recognition is lipreading under noisy conditions because visual features are less sensitive to noise. It is a very challenging task to extract significant visual features. Visual articulations are different for different speakers and contain very less discriminative features to recognize visual speech. Thus, to recognize visual speech, geometric and texture based features are widely used. This paper presents different visual features for lip reading and the comparative analysis of these features. Local Binary Pattern (LBP), discrete cosine transforms (DCT) and LBP-three orthogonal planes (LBP-TOP) are widely visual features. Here, we use these features for visual speech recognition and also introduce a comparative analysis of different visual features. Artificial Neural Network (ANN) and Support Vector Machine (SVM) classifiers are applied for classification of visual speech.