· 3월 27일 PM 04:01
Current sign language AI focuses on hand/body tracking but misses the linguistic complexity that makes sign languages full natural languages.
Most sign language AI systems treat the problem as pose estimation + classification. Track hand positions, match to a dictionary of signs. But sign languages are far more complex — they use simultaneous channels (both hands, facial expressions, body posture, spatial reference), and the meaning changes with context, speed, and spatial relationships. We're building our system on multimodal transformers that process all channels simultaneously.
You're identifying a critical gap. The pose-estimation approach treats sign language like a cipher for spoken language — each sign mapping to a word. But linguistically, sign languages are full natural languages with their own grammar, syntax, and pragmatics.
For example, in ASL, the sentence 'I give you a book' and 'You give me a book' use the same hand shape but different spatial trajectories. The spatial relationship IS the grammar. A pose-classification system that ignores spatial dynamics would confuse these fundamentally different meanings.
Multimodal transformers that attend across channels simultaneously is the right approach. The key challenge is training data — you need large-scale datasets annotated at the linguistic level, not just the pose level.
Data is our biggest bottleneck. We're working with Deaf community partners to build annotation tools that are accessible and respect the cultural context of Deaf communication. It's slower than scraping videos, but the quality difference is enormous.
Community-centered data collection is both ethically necessary and technically superior for this problem. The Deaf community has historically been excluded from the design of technology meant for them — 'nothing about us without us' should be the default.
The quality advantage is real: linguistic annotation by native signers captures nuances that non-signers would miss — like the difference between a raised eyebrow as a grammatical marker (yes/no question) versus an emotional expression. These distinctions are invisible to outside annotators.
This approach also builds trust and adoption. Technology for the Deaf community has a mixed history — many well-intentioned products failed because they were designed without community input. Your collaborative approach makes the product better and builds the social capital needed for adoption.