Major Features:
- Complete OpenSeeFace integration with bundled components and model downloading
- Sophisticated multi-point MAR calculation using both inner and outer mouth landmarks
- Unified preview video creation system for all providers
- Node renamed from "🎥 Mouth Movement Analyzer" to "🗣️ Silent Speech Analyzer"
OpenSeeFace Implementation:
- Handle 3D landmarks (X,Y,Z) with coordinate transformation for 68-point dlib model
- Fallback logic for missing inner mouth landmarks in bundled model
- Adaptive MAR calculation: outer mouth for stability, inner mouth for precision
- Enhanced threshold mapping (0.2-0.8 range) for better speech detection
- Automatic model downloading integrated with existing F5-TTS infrastructure
Preview System Enhancements:
- Unified PreviewVideoCreator utility for consistent preview generation
- Fast WebM encoding with optimized settings (CRF 45, speed 8)
- Full frame rate capture (removed frame skipping) for smoother previews
- Proper coordinate visualization for OpenSeeFace landmarks
UI/UX Improvements:
- Updated tooltips to mark OpenSeeFace as EXPERIMENTAL
- Enhanced provider descriptions with clear recommendations
- Better error handling and user feedback
Technical Fixes:
- Fix OpenSeeFace tracker RetinaFace conditional initialization
- Handle bundled vs downloaded model scenarios
- Proper SRT formatting for viseme sequences (fix TypeError with nested lists)
- Robust landmark attribute detection across different OpenSeeFace versions
diodiogod
bca3fa73fd