Pengenalan Viseme Dinamis Bahasa Indonesia Menggunakan Convolutional Neural Network

There has been very little researches on automatic lip reading in Indonesian language, especially the ones based on dynamic visemes. To improve the accuracy of a recognition process, for certain problems, choosing suitable classifiers or combining of some methods may be required. This study aims to classify five dynamic visemes of Indonesian language using a CNN (Convolutional Neural Network) and to compare the results with an MLP (Multi Layer Perceptron). Varying some parameters theoretically improving the recognition accuracy was attempted to obtain the best result. The data includes videos on pronunciation of daily words in Indonesian language by 28 subjects recorded in frontal view. The best recognition result gives 96.44% of validation accuracy using the CNN classifier with three convolution layers.