Readme for supplementary material for the Eurographics 2020 article "Style-Controllable Speech-Driven Gesture Synthesis Using Normalising Flows". 


These folders contain down-scaled (320x240) versions of all the videos included in our subjective evaluations. The original resolution shown to the subjects was 960x720.

The style-controlled videos comes in sequences of 3, were indices 0,3,6,... 1,4,7,... and 2,5,8... use the 15%, 50% and 85% percentile control values respectively.

Bottom-line (BL) stimuli are included for each evaluation.

The sub-folder 'other_speakers' contains extra material showing how a model trained on the same data generalises to audio from unknown speakers not in the training set.

The 'erratum-folder' describes a bug in one of our style-controls found shortly before submitting the camera-ready version.

We also provide a video presenting our results.

Thanks!