Leitura labial em vídeos com a utilização de aprendizado profundo

Abstract

In recent years, the evolution and popularization of machine learning techniques have allowed sophisticated automated lip reading algorithms to be proposed. To accompany such algorithms, datasets composed of images prepared for this purpose were created, which may contain images related to letters, digits, words or phrases spoken by one or more speakers. Thus, the present work shows the construction of a dataset composed of images related to words being spoken by a single speaker. In addition, deep learning models based on convolutional neural networks was also used to perform automated lip reading on this dataset reaching up to 90.28% accuracy.