Fake it to make it: Using synthetic data to remedy the data shortage in joint multimodal speech-and-gesture synthesis
Shivam Mehta, Anna Deichler, Jim O'Regan, Birger Moëll, Jonas Beskow, Gustav Eje Henter, Simon Alexanderson
TL;DR
This work tackles the data scarcity barrier in joint multimodal speech-and-gesture synthesis by leveraging a chain of unimodal synthetic data generators to pre-train a unified model. The authors introduce MAGI, an enhanced architecture that adds probabilistic duration modelling, explicit pitch/energy prosody predictors, and multispeaker capability, building on Match-TTSG with a refined information-flow design. Empirical results show that pre-training on large synthetic corpora improves both speech naturalness and gesture realism, with the architectural advances providing additional gains when such data is available; however, cross-modal appropriateness continues to lag behind human performance. The approach demonstrates a scalable pathway to high-quality, controllable multimodal synthesis using synthetic data, with practical implications for virtual agents and embodied conversational systems.
Abstract
Although humans engaged in face-to-face conversation simultaneously communicate both verbally and non-verbally, methods for joint and unified synthesis of speech audio and co-speech 3D gesture motion from text are a new and emerging field. These technologies hold great promise for more human-like, efficient, expressive, and robust synthetic communication, but are currently held back by the lack of suitably large datasets, as existing methods are trained on parallel data from all constituent modalities. Inspired by student-teacher methods, we propose a straightforward solution to the data shortage, by simply synthesising additional training material. Specifically, we use unimodal synthesis models trained on large datasets to create multimodal (but synthetic) parallel training data, and then pre-train a joint synthesis model on that material. In addition, we propose a new synthesis architecture that adds better and more controllable prosody modelling to the state-of-the-art method in the field. Our results confirm that pre-training on large amounts of synthetic data improves the quality of both the speech and the motion synthesised by the multimodal model, with the proposed architecture yielding further benefits when pre-trained on the synthetic data. See https://shivammehta25.github.io/MAGI/ for example output.