GiantMIDI-Piano [1] is a classical piano MIDI dataset contains 10,855 MIDI files of 2,786 composers. The curated subset by constraining composer surnames contains 7,236 MIDI files of 1,787 composers. GiantMIDI-Piano are transcribed from live recordings with a high-resolution piano transcription system [2].
Here is the demo of GiantMIDI-Piano: https://www.youtube.com/watch?v=5U-WL0QvKCg
Transcribed MIDI files of GiantMIDI-Piano can be viewed at midis_preview directory.
Follow disclaimer.md to agree a disclaimer and download a stable version of GiantMIDI-Piano (193 MB).
Users can acquire GiantMIDI-Piano by downloading all audio recordings, and transcribing them into MIDI files following the rest part of this repo. The transcription takes ~200 hours on a single GPU card.
Install PyTorch (>=1.4) following https://pytorch.org/.
The above links also include a curated subset. The curated subset constrains the YouTube titles should contain composers surnames.
pip install -r requirements.txt
Download audio recordings from YouTube using the following scripts. Approximately 10,855 audio recordings can be downloaded. There can be audios no longer downloadable.
WORKSPACE="./workspace"
mkdir -p $WORKSPACE
cp "resources/full_music_pieces_youtube_similarity_pianosoloprob.csv" $WORKSPACE/"full_music_pieces_youtube_similarity_pianosoloprob.csv"
# Download all mp3s. Users could split the downloading into parts to speed up the downloading. E.g.,
python3 dataset.py download_youtube_piano_solo --workspace=$WORKSPACE --begin_index=0 --end_index=30000
python3 dataset.py download_youtube_piano_solo --workspace=$WORKSPACE --begin_index=30000 --end_index=60000
python3 dataset.py download_youtube_piano_solo --workspace=$WORKSPACE --begin_index=60000 --end_index=90000
python3 dataset.py download_youtube_piano_solo --workspace=$WORKSPACE --begin_index=90000 --end_index=120000
python3 dataset.py download_youtube_piano_solo --workspace=$WORKSPACE --begin_index=12000 --end_index=150000
The downloaded mp3 files look like:
mp3s_piano_solo (10,855 files) ├── Aaron, Michael, Piano Course, V8WvKK-1b2c.mp3 ├── Aarons, Alfred E., Brother Bill, Giet2Krl6Ww.mp3 └── ...
# Transcribe all mp3s to midi files. Users could split the transcription into parts to speed up the transcription. E.g.,
python3 audios_to_midis.py transcribe_piano --workspace=$WORKSPACE --mp3s_dir=$WORKSPACE"/mp3s_piano_solo" --midis_dir=$WORKSPACE"/midis" --begin_ind=0 --end_index=30000
python3 audios_to_midis.py transcribe_piano --workspace=$WORKSPACE --mp3s_dir=$WORKSPACE"/mp3s_piano_solo" --midis_dir=$WORKSPACE"/midis" --begin_ind=30000 --end_index=60000
python3 audios_to_midis.py transcribe_piano --workspace=$WORKSPACE --mp3s_dir=$WORKSPACE"/mp3s_piano_solo" --midis_dir=$WORKSPACE"/midis" --begin_ind=60000 --end_index=90000
python3 audios_to_midis.py transcribe_piano --workspace=$WORKSPACE --mp3s_dir=$WORKSPACE"/mp3s_piano_solo" --midis_dir=$WORKSPACE"/midis" --begin_ind=90000 --end_index=120000
python3 audios_to_midis.py transcribe_piano --workspace=$WORKSPACE --mp3s_dir=$WORKSPACE"/mp3s_piano_solo" --midis_dir=$WORKSPACE"/midis" --begin_ind=120000 --end_index=150000
The transcribed MIDI files look like:
midis (10,855 files) ├── Aaron, Michael, Piano Course, V8WvKK-1b2c.mid ├── Abel, Frederic, Lola Polka, SLNJF0uiqRw.mid └── ...
The transcription of all audio recordings may take around 10 days on a single GPU card.
Details of scripts can be viewed at scripts
All statistics and figures in [1] can be reproduced by:
./scripts/3_statistics.shIf users met "Too many requests! Sleep for 3600 s" when downloading, it means that YouTube has limited the number of videos for downloading. Users could either 1) Wait until YouTube unblock your IP (1 days or a few weeks), or 2) try to use another machine with a different IP for downloading.
Qiuqiang Kong, qiuqiangkong@gmail.com
[1] Qiuqiang Kong, Bochen Li, Jitong Chen, and Yuxuan Wang. "GiantMIDI-Piano: A large-scale MIDI dataset for classical piano music." arXiv preprint arXiv:2010.07061 (2020). https://arxiv.org/pdf/2010.07061
CC BY 4.0