An Evaluation Model For Speech-Driven Gesture Synthesis

The research and development of embodied agents with advanced relational capabilities is constantly evolving. In recent years, the development of behavioural signal generation models to be integrated in social robots and virtual characters, is moving from rule-based to data-driven approaches, requiring appropriate and reliable evaluation techniques. This work proposes a novel machine learning approach for the evaluation of speech-to-gestures models that is independent from the audio source. This approach enables the measurement of the quality of gestures produced by these models and provides a benchmark for their evaluation. Results show that the proposed approach is consistent with evaluations made through user studies and, furthermore, that its use allows for a reliable comparison of speech-to-gestures state-of-the-art models.