Music2MIDI: Pop Music to MIDI Piano Cover Generation

Music2MIDI: a deep learning model that generates MIDI piano cover songs from the original song audio. Official repository for Music2MIDI: Pop Music to MIDI Piano Cover Generation.

Setup

FFmpeg

Install FFmpeg.

FluidSynth

Install FluidSynth.

Conda Environment

conda env create -f environment.yaml
conda activate music2midi

Model Checkpoint

Download trained model checkpoint in Releases.

Usage

Web UI

python webui.py --ckpt [ckpt_path]

Python API

See demo.ipynb.

Dataset preparation

Download the dataset in Releases then unzip it.

Download the audio files from YouTube.

python data/download_youtube.py [data_dir]

Data preprocessing

The released dataset is already preprocessed. See dataset.md if you want to prepare your own dataset.

Training

python train.py [data_dir]

Evaluation

Evaluate the model based on melody chroma accuracy.

python evaluate.py [data_dir] --ckpt [ckpt_path]

Troubleshooting

• If you encounter the error module 'soundfile' has no attribute 'SoundFileRuntimeError', uninstall soundfile then install it again:

  pip uninstall pysoundfile
  pip uninstall soundfile
  pip install soundfile
  

• If you encounter problems with FluidSynth like /usr/lib/libinstpatch-1.0.so.2: undefined symbol: g_once_init_leave_pointer, consider downgrading libinstpatch in your system.