LLM-based Chinese Art Song Assisted-Composition Agent System

Pipeline

1. Emotion analysis — an LLM maps the user description to musical parameters (title, key, time signature, tempo, mood).
2. Melody generation — the fine-tuned NotaGen model generates an ABC melody conditioned on the key, meter, and tempo.
3. Lyric writing — musical-form analysis, outline planning, phrase-by-phrase lyric generation, followed by syllable-count, rhyme, and semantic correction.
4. Conversion — LLM-based lyric–note alignment, then export to MIDI and MusicXML.

Requirements

• Python 3.10
• pip install -r requirements.txt
• For a CUDA 12.1 build of PyTorch: pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cu121
• The external tool abc2midi (from the abcMIDI package), installed at the system level (macOS: brew install abcmidi; Debian/Ubuntu: apt install abcmidi).

Configuration

Melody model weights

Usage

1. Install the Python dependencies and the abc2midi tool as described above.
2. Place the NotaGen melody weights at the default path, or specify their location:

1. Set the LLM environment variables in the Configuration section.
2. Run the system with a natural-language description:

1. Find the generated ABC, MIDI, and MusicXML files in output/.

Layout

config.py            Global configuration (LLM access, paths, generation limits)
main.py              Command-line entry point
pipeline.py          Four-stage composition pipeline
modules/             Emotion analysis, lyric generation, NotaGen interface, conversion
melody/              NotaGen melody model: data, pretrain, finetune, RL, inference, generate.py
requirements.txt     Python dependencies
run.sh               Convenience launcher