Reconstructing intelligible audio speech from visual speech features

This work describes an investigation into the feasibility of producing intelligible audio speech from only visual speech features. The proposed method aims to estimate a spectral envelope from visual features which is then combined with an artificial excitation signal and used within a model of speech production to reconstruct an audio signal. Different combinations of audio and visual features are considered, along with both a statistical method of estimation and a deep neural network. The intelligibility of the reconstructed audio speech is measured by human listeners, and then compared to the intelligibility of the video signal only and when combined with the reconstructed audio.