Improving Viseme Recognition Using GAN-Based Frontal View Mapping

Deep learning methods have become the standard for Visual Speech Recognition problems due to their high accuracy results reported in the literature. However, while successful works have been reported for words and sentences, recognizing shorter segments of speech, like phones, has proven to be much more challenging due to the lack of temporal and contextual information. Also, head-pose variation remains a known issue for facial analysis with direct impact in this problem. In this context, we propose a novel methodology to tackle the problem of recognizing visemes – the visual equivalent of phonemes – using a GAN to artificially lock the face view into a perfect frontal view, reducing the view angle variability and simplifying the recognition task performed by our classification CNN. The GAN is trained using a large-scale synthetic 2D dataset based on realistic 3D facial models, automatically labelled for different visemes, mapping a slightly random view to a perfect frontal view. We evaluate our method using the GRID corpus, which was processed to extract viseme images and their corresponding synthetic frontal views to be further classified by our CNN model. Our results demonstrate that the additional synthetic frontal view is able to improve accuracy in 5.9% when compared with classification using the original image only.