mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-09-01 19:42:03 +00:00
Vance Ingalls
37404f23da
## Summary **Whisper improvements:** - Auto-detect language and switch from `.en` to multilingual model when needed - Detect speech onset in WAV to strip hallucinated words before speech begins - Merge whisper-cpp token fragments: contractions (`didn` + `'t` → `didn't`), split capitals (`C` + `aught` → `Caught`), dropped-g (`shin` + `in'` → `shinin'`) - Interpolate zero-duration word clusters for reliable karaoke timing **Captions skill updates (folded from #176):** - Rewrite script-to-style mapping as an energy detection table (high → low) with mandatory animation requirements: karaoke baseline, 2+ highlight techniques, kinetic exits - Replace `tl.call()` per-frame audio-reactive pattern with group-level GSAP tweens — read peak bass/treble for each group's time range and modulate entrance intensity at build time - Add transcript quality check with automatic retry rules (>20% music tokens = retry with larger model) - Add caption word structure lint rule (`.caption-group` + `<span>`) for studio editor compatibility **Multilingual defaults (folded from #186):** - Default whisper model changed from `small.en` to `small` to prevent silent translation of non-English audio - Added non-negotiable language rule to captions skill ## Test plan - [ ] `pnpm test` passes (contraction merging, fragment merging, zero-duration interpolation, speech onset) - [ ] Transcribe non-English audio — verify it transcribes in original language, not translates - [ ] Skill files render correctly, cross-references resolve - [ ] `dynamic-techniques.md` audio-reactive section uses `tl.to()`/`tl.set()` only, no `tl.call()` loops 🤖 Generated with [Claude Code](https://claude.com/claude-code)