Enhancement of VSR using low dimension visual feature

This paper presents a study about the low dimension visual (LDV) space features and investigates the improvement in audio visual automatic speech recognition using different set of visual features. The experiment is divided into three sub-sections; in first phase the recognition is performed on 12 static DCT features; in second phase the recognition is performed for combination of 6 static and 6 dynamic features and in third phase the recognition is performed on 12 low dimension DCT feature. For this research work Hindi AMUAV (Aligarh Muslim University Audio-Visual) database was developed in which audio sample at 44.1 kHz and video sample at 25 frames per second was opted. Hidden Markov Model (HMM) tool kit with left-right HMMs modeled was used for recognition and an overall improvement of 26.04% in word recognition is achieved with LDV space features.