Human-Centered Editable Speech-to-Sign-Language Generation via Streaming Conformer-Transformer and Resampling Hook

User Control Improves Accessibility Outcomes: The system's editable JSON intermediate representation—allowing deaf users and sign language experts to inspect and modify individual sign segments—resulted in a 13-point improvement in System Usability Scale (SUS) scores and a 6.7-point reduction in cognitive load compared to baseline systems, demonstrating that transparency and control are critical for accessibility technology.

Human-in-the-Loop Feedback Enhances AI Performance: By incorporating user edits and ratings into a continuous model adaptation loop, the system achieved significant improvements in naturalness and trust (p < .001) in studies with 20 deaf signers and 5 professional interpreters, showing that AI systems serving marginalized communities benefit from iterative refinement based on user input rather than one-time training.

Real-Time Performance Enables Practical Deployment: The system achieves 13 ms average frame-inference time and 103 ms end-to-end latency on consumer hardware (RTX 4070), making it feasible for real-world applications—a crucial consideration for accessibility tools that must function reliably in everyday communication scenarios.

Explainability Builds Trust in Accessibility AI: The transparent, editable design allows users to understand and modify the AI's outputs at a granular level, which the study found significantly increased trust among deaf signers and professional interpreters—highlighting that accessibility technology must prioritize interpretability, not just accuracy.