This codebase automates the collection of fine-aligned audio-MIDI data using (e.g.) a Yamaha Disklavier self-playing piano.
You can download a 16 kHz downsampled monophonic mixture of a re-performance
of the MAESTRO dataset with piano_capture on Zenodo (Studio MAESTRO: https://zenodo.org/records/10082144).
Recommended installation:
pip install -e .
Example usage:
INPUT_MIDI_ROOT=/path/to/midi_root
OUTPUT_AUDIO_ROOT=/path/to/audio_root
MIDI_OUT_BUS="YAMAHA USB Device Port"
INPUT_AUDIO_DEVICE=0
python -m piano_capture.app \
$INPUT_MIDI_ROOT \
$OUTPUT_AUDIO_ROOT \
$MIDI_OUT_BUS \
$INPUT_AUDIO_DEVICE \
--sample_rate=44100 \
--num_channels=2 \
--output_suffix="_v1" \
--delay-start
See python -m piano_capture.app --help for further information.
The program will load all MIDI files beneath the specified path and perform each on the Disklavier one-by-one. One audio file will be created per input MIDI file. The hierarchy created beneath the output directory will match that of the input MIDI directory.
Why would anyone use this?
Excellent question! Perhaps you work at a university or research lab focused on music informatics. And perhaps you have access to a MIDI-acoustic piano, such as the Yamaha Disklavier. And perhaps you want to create a dataset of piano audio aligned with MIDI, in order to train automatic piano transcription systems!
During a research collaboration with Yamaha Music Research, I used this code to capture nearly 500 hours of aligned piano data!
'MemoryError: Cannot allocate write+execute memory for ffi.callback()'
spatialaudio/python-sounddevice#397
error: legacy-install-failure
× Encountered error while trying to install package.
╰─> python-rtmidi
For newer versions of Python, you may need to remove the python-rtmidi dependency and install a manual
pip install git+https://github.com/SpotlightKid/python-rtmidi.git@eb16ab3268b29b94cd2baa6bfc777f5cf5f908ba#egg=python-rtmidi
(See this issue in the python-rtmidi project for more info.)