Abstract Method Quantitative Results Acknowledgements BibTeX
BMVC 2026

M3T: Discrete Multi-Modal Motion Tokens for Sign Language Production

Alexandre Symeonidis-Herzig1,  Jianhe Low1,  Ozge Mercanoglu Sincan1,  Richard Bowden1

1CVSSP, University of Surrey, United Kingdom

Paper arXiv Code Coming soon
M3T architecture overview

Fig. 1 — M3T architecture: text and multi-modal motion tokens are combined in a shared vocabulary and processed by an autoregressive transformer for both Sign Language Production (SLP) and Sign Language Translation (SLT).


Abstract

Sign language production requires more than hand motion generation. Non-manual features, including mouthings, eyebrow raises, gaze, and head movements, are grammatically obligatory and cannot be recovered from manual articulators alone. Existing 3D production systems face two barriers to integrating them: the standard body model provides a facial space too low-dimensional to encode these articulations, and when richer representations are adopted, standard discrete tokenization suffers from codebook collapse, leaving most of the expression space unreachable. We propose SMPL-FX, which couples FLAME's rich expression space with the SMPL-X body, and tokenize the resulting representation with modality-specific Finite Scalar Quantization (FSQ) VAEs for body, hands, and face. M3T is an autoregressive transformer trained on this multi-modal motion vocabulary, with an auxiliary translation objective that encourages semantically grounded embeddings. Across three standard benchmarks (How2Sign, CSL-Daily, Phoenix14T) M3T achieves state-of-the-art sign language production quality, and on NMFs-CSL, where signs are distinguishable only by non-manual features, reaches 58.3% accuracy against 49.0% for the strongest comparable pose baseline.


Method

SMPL-FX: Expressive Full-Body Representation

FLAME Expression Space Integrates FLAME's high-dimensional facial model into SMPL-X, providing the rich expression space needed to encode mouthings, eyebrow raises, gaze, and head movements.
Unified Body–Face Model SMPL-FX couples expressive facial articulations with the full-body SMPL-X skeleton, enabling joint modelling of manual and non-manual features in a single representation.
Beyond Standard Body Models Standard SMPL-X facial PCA space is too low-dimensional to encode grammatically obligatory non-manual features — SMPL-FX resolves this limitation.

Modality-Specific FSQ VAEs: Discrete Token Learning

Finite Scalar Quantization Replaces standard VQ codebooks with FSQ, which deterministically maps each dimension to a finite set of scalar values, eliminating codebook collapse entirely.
Three Modality-Specific VAEs Separate VAEs for body, hands, and face allow each modality to be tokenized at its own appropriate resolution and codebook size.
No Codebook Collapse Standard VQ tokenization leaves most of the expression space unreachable when applied to facial articulations — FSQ guarantees full utilisation of the token vocabulary.

M3T: Autoregressive Multi-Modal Transformer

Shared Multi-Modal Vocabulary Text tokens and multi-modal motion tokens (body, hands, face) are combined into a single vocabulary, enabling the model to reason jointly over language and motion.
Auxiliary Translation Objective An auxiliary Sign Language Translation (SLT) task is trained alongside production, encouraging semantically grounded token embeddings that better capture sign meaning.
Language Tag Conditioning Language tags in the vocabulary support conditioning on different sign language corpora within a unified model.
Dual Task: SLP & SLT The same architecture handles both Sign Language Production (text → motion) and Sign Language Translation (motion → text), with shared representations benefiting both directions.

Quantitative Results

58.3%
Top-1 accuracy on the NMFs-CSL Confusing split — signs distinguishable only by non-manual features. M3T reaches 58.3% versus 49.0% for BEST, the strongest directly comparable pose baseline, demonstrating that expressive facial modelling is critical for sign language production.

Sign Language Production — Standard Benchmarks

Method How2Sign CSL-Daily Phoenix14T
DTW-PA-JPE ↓ DTW-JPE ↓ DTW-PA-JPE ↓ DTW-JPE ↓ DTW-PA-JPE ↓ DTW-JPE ↓
BodyHandBodyHand BodyHandBodyHand BodyHandBodyHand
Prior Work
NAR13.9411.80
Prog. Trans.14.1511.5714.7430.1715.9812.9116.3032.6313.6711.9515.0131.77
Text2Mesh13.9913.4715.5032.9713.4712.1013.7630.3713.4812.0614.0431.64
Adv. Train.13.7811.17
T2S-GPT11.486.3912.6518.4411.945.9312.3215.4310.386.4711.6519.09
NSA7.837.33
S-MotionGPT11.234.3912.4113.7410.813.7811.5811.319.453.4110.429.08
SOKE6.822.357.7510.086.241.717.389.684.771.386.047.72
M3T (Ours)
M3T w/o SLT4.522.425.028.944.131.594.779.943.541.544.179.14
M3T w/ SLT4.282.284.869.443.971.514.639.283.411.464.028.55

Comparison with state-of-the-art SLP (text-to-sign) methods. DTW-PA-JPE measures pose-aligned joint error after rigid normalisation, while DTW-JPE measures raw joint position error; both are reported separately for body and hand joints. Bold marks the best value in each column. M3T reduces body DTW-PA-JPE by 37–29% relative to the strongest prior method (SOKE) across all three benchmarks, with the auxiliary SLT objective improving 9 of 12 metrics over the ablated variant.

Non-Manual Feature Evaluation — NMFs-CSL

Method Overall Confusing Normal
T-1 ↑T-5 ↑ T-1 ↑T-5 ↑ T-1 ↑T-5 ↑
RGB-Based Methods
3D-R5062.182.943.172.487.497.0
DNF55.882.451.971.486.397.0
I3D64.488.047.381.887.197.3
TSM64.588.742.981.093.399.0
SlowFast66.386.647.077.492.098.9
GLE-Net69.088.150.679.693.699.3
HMA64.791.042.384.894.699.3
NLASLR83.198.3
Pose-Based Methods
ST-GCN59.986.842.279.483.496.7
SignBERT67.095.346.492.194.599.6
BEST68.594.449.090.394.699.7
ScalingUp80.297.567.295.797.599.8
M3T74.195.158.393.093.597.6

SLR results on NMFs-CSL, separated by RGB and pose modalities. denotes pre-training on large sign datasets. On the Confusing split — signs distinguishable only by non-manual features — M3T reaches 58.3% Top-1 against 49.0% for BEST, the strongest directly comparable pose baseline, a +9.3 pp gain.


Acknowledgements

This work was supported by EPSRC grant APP24554 (SignGPT — EP/Z535370/1), EPSRC grant APP78083 (UMCS — UKRI3927), and through funding from Google.org via the AI for Global Goals scheme. The authors acknowledge the use of Isambard-AI National AI Research Resource (AIRR), funded by UK DSIT via UKRI and STFC [ST/AIRR/I-A-I/1023]. This work reflects only the authors' views; the funders are not responsible for any use that may be made of the information it contains.


BibTeX

@article{symeonidisherzig2026m3t,
  title   = {M3T: Discrete Multi-Modal Motion Tokens for
             Sign Language Production},
  author  = {Symeonidis-Herzig, Alexandre and Low, Jianhe and
             Sincan, Ozge Mercanoglu and Bowden, Richard},
  journal = {arXiv preprint arXiv:2603.23617},
  year    = {2026},
  eprint  = {2603.23617},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi     = {10.48550/arXiv.2603.23617},
}