Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unified speech and gesture synthesis using flow matching

Shivam Mehta, Ruibo Tu, Simon Alexanderson, Jonas Beskow, Éva Székely, Gustav Eje Henter

TL;DR

The paper tackles joint generation of speech acoustics and co-speech gestures from text by introducing a truly unified multimodal architecture trained with optimal-transport conditional flow matching (OT-CFM). It replaces separate diffusion pathways with a single 1D Transformer U-Net decoder and models the joint distribution $p( ext{acoustics}, ext{motion} | ext{text})$ instead of independent marginals, yielding improved cross-modal coherence. OT-CFM enables sampling with very few ODE steps, delivering substantial speedups and reduced memory compared to prior state-of-the-art Diff-TTSG, while achieving higher subjective and objective quality (e.g., lower WER and higher MOS/MAS). The results imply that integrated, flow-matched multimodal synthesis is both feasible and advantageous for realistic, synchronous speech and gesture generation in interactive agents and XR applications.

Abstract

As text-to-speech technologies achieve remarkable naturalness in read-aloud tasks, there is growing interest in multimodal synthesis of verbal and non-verbal communicative behaviour, such as spontaneous speech and associated body gestures. This paper presents a novel, unified architecture for jointly synthesising speech acoustics and skeleton-based 3D gesture motion from text, trained using optimal-transport conditional flow matching (OT-CFM). The proposed architecture is simpler than the previous state of the art, has a smaller memory footprint, and can capture the joint distribution of speech and gestures, generating both modalities together in one single process. The new training regime, meanwhile, enables better synthesis quality in much fewer steps (network evaluations) than before. Uni- and multimodal subjective tests demonstrate improved speech naturalness, gesture human-likeness, and cross-modal appropriateness compared to existing benchmarks. Please see https://shivammehta25.github.io/Match-TTSG/ for video examples and code.

Unified speech and gesture synthesis using flow matching

TL;DR

The paper tackles joint generation of speech acoustics and co-speech gestures from text by introducing a truly unified multimodal architecture trained with optimal-transport conditional flow matching (OT-CFM). It replaces separate diffusion pathways with a single 1D Transformer U-Net decoder and models the joint distribution instead of independent marginals, yielding improved cross-modal coherence. OT-CFM enables sampling with very few ODE steps, delivering substantial speedups and reduced memory compared to prior state-of-the-art Diff-TTSG, while achieving higher subjective and objective quality (e.g., lower WER and higher MOS/MAS). The results imply that integrated, flow-matched multimodal synthesis is both feasible and advantageous for realistic, synchronous speech and gesture generation in interactive agents and XR applications.

Abstract

As text-to-speech technologies achieve remarkable naturalness in read-aloud tasks, there is growing interest in multimodal synthesis of verbal and non-verbal communicative behaviour, such as spontaneous speech and associated body gestures. This paper presents a novel, unified architecture for jointly synthesising speech acoustics and skeleton-based 3D gesture motion from text, trained using optimal-transport conditional flow matching (OT-CFM). The proposed architecture is simpler than the previous state of the art, has a smaller memory footprint, and can capture the joint distribution of speech and gestures, generating both modalities together in one single process. The new training regime, meanwhile, enables better synthesis quality in much fewer steps (network evaluations) than before. Uni- and multimodal subjective tests demonstrate improved speech naturalness, gesture human-likeness, and cross-modal appropriateness compared to existing benchmarks. Please see https://shivammehta25.github.io/Match-TTSG/ for video examples and code.
Paper Structure (9 sections, 1 equation, 1 figure, 2 tables)

This paper contains 9 sections, 1 equation, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Prior work
  4. Conditional flow matching
  5. Method
  6. Experiments
  7. Data and systems
  8. Evaluation, results, and discussion
  9. Conclusions and future work

Figures (1)

  • Figure 1: Schematic overview of Match-TTSG training and synthesis. Still frames of the avatar used to visualise motion can be seen top right.