Spaces:
Running
Running
Commit History
-
Use sentence-based window sliding instead of word-based
- Fix complete sentences instead of individual words - Match Python implementation behavior (STT/smart_progressive_streaming.py) - Apply to both transcribeIncremental and transcribeBatch methods - Improves transcription quality by preserving sentence boundaries Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>c9d9124 -
Make UI more minimal and compact
- Remove "About This Demo" section for cleaner interface - Make controls thinner with reduced padding (p-6 → p-4) - Reduce label sizes (text-sm → text-xs) and spacing - Ensure all boxes (controls, transcription, metrics) are aligned with consistent max-width Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>b76aacc
-
Optimize performance with webgpu-hybrid and fix RTF display
- Use webgpu-hybrid backend (FP32 encoder on WebGPU + INT8 decoder on WASM) - Achieve RTF 25x+ (1.8x faster than parakeet.js demo) - Fix RTF calculation to show speed factor (higher is better) - Extend progressive streaming window to 15 seconds - Update UI: rename 'Audio Duration' to 'Window Size', remove 'Update Rate' - Performance: ~230ms for 5.8s audio (vs 411ms in demo) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>ea6aed8 -
Final optimizations: VAD, 5s window, word-level sliding
- Simple amplitude-based VAD skips transcription during silence - Reduced window from 15s to 5s for faster sliding - Word-level fallback ensures window always slides at 5s - Removed excessive debug logging - Working with SharedArrayBuffer (12 threads) on HF Spaces! Performance: RTF 2-4x with encode times 400-1300ms (vs 5000ms single-threaded) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>09b5ac6 -
Try COEP credentialless without COOP for SharedArrayBuffer
Using only COEP: credentialless should enable SharedArrayBuffer without breaking iframe embedding (no COOP required). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>0e8835b -
Add cross-origin headers and optimize progressive streaming
- Add _headers file with COEP/COOP for WebGPU support on HF Spaces - Implement simple amplitude-based VAD to skip transcription during silence - Reduce max window size from 15s to 5s for faster processing - Add word-level fallback when no sentence boundaries detected - Remove excessive debug logging for cleaner console output Fixes model loading hang on deployed Space due to missing cross-origin headers. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>e74c555 -
Restructure for HuggingFace Space compatibility
- Move built files (index.html, assets/) to root for demo to work - Move source code to source/ folder for visibility - Clean file organization: demo works + source is browsable Structure: - Root: index.html, assets/ (working demo) - source/: all source code, configs, docs - README.md: Space documentation Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>0f739b8