Robust Visual Voice Activity Detection Using Long Short-Term Memory Recurrent Neural Network

摘要

Many traditional visual voice activity detection systems utilize features extracted from mouth region images which are sensitive to noisy observations of the visual domain. In addition, hyperparameters of the feature extraction process modulating the desired compromise between robustness, efficiency, and accuracy of the algorithm are difficult to be determined. Therefore, a visual voice activity detection algorithm which only utilizes simple lip shape information as features and a Long Short-Term Memory recurrent neural network (LSTM-RNN) as a classifier is proposed. Face detection is performed by structural SVM based on histogram of oriented gradient (HOG) features. Detected face template is used to initialize a kernelized correlation filter tracker. Facial landmark coordinates are then extracted from the tracked face. Cen-troid distance function is applied to the geometrically normalized landmarks surrounding the outer and inner lip contours. Finally, discriminative (LSTM-RNN) and generative (Hidden Markov Model) methods are used to model the temporal lip shape sequences during speech and non-speech intervals and their classification performances axe compared. Experimental results show that the proposed algorithm using LSTM-RNN can achieve a classification rate of 98% in labeling speech and non-speech periods. It is robust and efficient for realtime applications.