Skarly is a local-first research prototype for vocal-to-music and music-to-music generation. It runs as a guest studio with no account system and no cloud services.
- Expo / React Native web interface in
lyricmorph-mobile/ - FastAPI backend in
lyricmorph-backend/ - SQLite project metadata and local filesystem audio storage
- ACE-Step 1.5 for five AI backing versions
- Demucs for vocal isolation from full-song inputs
- FFmpeg for decoding, normalization, mixing and exports
- Whisper, Basic Pitch and a reviewed local audio-intelligence checkpoint for analysis
This release is deliberately limited to a local guest session, local persistence, an in-process worker, and the checked-in generation stack.
- Executed notebook:
research/Skarly_Audio_Intelligence_Research.ipynb - Reproduction notes:
research/README.md - UI catalogue:
docs/ui-screenshots/README.md - PDF report:
output/pdf/Skarly_Research_Internship_Report.pdf - Editable report:
docs/research-report/Skarly_Research_Internship_Report.docx
See QUICKSTART.md. The normal stack is:
- Start the ACE-Step API with
tools/start-ace-step-api.ps1. - Start FastAPI with
tools/start-local-studio.ps1. - Run
npm installandnpm run webfromlyricmorph-mobile/. - Open the Expo URL, enter the Local Guest Studio, and upload or record audio.
The ACE-Step repository and model weights are installed separately under a sibling skarly-ai-repos/ACE-Step-1.5 checkout. Generated audio, downloaded datasets, dependencies, caches and local environment files are intentionally excluded from Git.
Skarly is a private local research tool. It does not provide payments, public feeds, voice cloning, public remix distribution or account-based synchronization. Only use audio you have permission to process.