Sign language machine translation over 40 years: from HMMs to LLMs and diffusion models
abenitezburraco · x · 2026-10-28
A survey charting four decades of sign language translation and production (SLT/SLP).
- 1980–2000 early approaches: Tamura et al.'s 1988 SLR, HMM-based SLR (1997).
- 2000–2015 feature-driven: rule-based generation, SVM SLR, linguistic corpora, Kinect datasets, random-forest feature selection.
- 2015–present deep learning: CNN-HMM hybrid CSLR (Koller et al.), neural SLT/SLP, OpenPose/MediaPipe pose estimation, Phoenix-2014T and CSL-Daily/How2Sign benchmarks, gloss-free translation (2023), diffusion-based sign production, and since 2024 LLM-based SLT with larger datasets like YT-SL-25.
The trajectory runs from handcrafted features to end-to-end neural models and now gloss-free LLM-era translation.
More from Research
- Where to find the colors your screen can't show you: a deep dive into gamuts and cones — moultano · 2026-10-28
- Claude designs 35,000 antibodies in a month as Anthropic pushes into drug discovery — minchoi · 2026-10-28
- Dev fine-tuned 1.2B-4B models to clean voice transcripts locally — beats cloud APIs — stephan273 · 2026-10-28
- Forecasting agent says Kalshi's 32% on P vs NP by 2030 is sentiment, not signal — rohanpaul_ai · 2026-10-28
- ColBERT deep dive: re-ranking can't recover the 18.6% of passages BM25 misses — antoine_chaffin · 2026-10-28
- AI Researcher on the Job Market Reflects on the Art of Problem Selection — simi_97k · 2026-10-28