MuScriptor
Audio-to-MIDI transcription using llama.cpp as the transformer backend.
8GB VRAM, 8x fast
Prerequisites
- Python with torch, torchaudio
- https://github.com/byzp/llama.cpp built with
llama-server or download from this repo(linux amd64 cu128)
- Model file:
model_fp32_8k.gguf conditioning_weights.safetensors
Extract Conditioning Weights
Only needed once, if starting from the full safetensors:
python extract_conditioning_weights.pyThis creates conditioning_weights.safetensors (~11 MB) from the full model weights.
This model can also be downloaded from this repo
CLI Transcription
Terminal 1 — start llama-server:
llama-server -m model_fp32_8k.gguf --port 8081 -ngl 99 --parallel 8Terminal 2 — run transcription:
python muscriptor_cli.py audio.mp3 -o out.mid
