n-Gram modeling of relevant features for lip-reading
Preety Singh, Vijay Laxmi, Manoj Singh Gaur · 2012
In this paper, relevant features for a visual speech recognition system are selected using Minimum Redundancy Maximum Relevance (mRMR) method. Feature vectors, with varying number of relevant attributes, are tested to determine the most optimal feature set. It is observed that a few relevant attributes perform considerably well compared to the complete feature vector. Using this feature set as a base vector, concatenation of features is done frame-wise to build n-gram models, so as to capture the temporal behaviour of speech. It is observed that the base mRMR feature vector is able to outperform the dynamic model of n-grams. This feature vector, consisting of only significant attributes, requires less processing time due to its small size. Storage requirements are also considerably reduced.