The article presents a joint audio-video approach towards speaker identity conversion, based on statistical methods originally introduced for voice conversion. Using the experimental data from the 3D BIWI Audiovisual corpus of Affective Communication, mapping functions are built between each two speakers in order to convert speaker-specific features: speech signal and 3D facial expressions. The results obtained by combining audio and visual features are compared to corresponding results from earlier approaches, while outlining the improvements brought by introducing dynamic features and exploiting prosodic features.

Index Terms: speaker identity conversion, gaussian mixture model, dynamic features, prosodic features