ASMD is a framework for installing, using and creating music multimodal datasets including (for now) audio and scores.
This is the repository for paper [1]
Read more in the docs.
- To install:
pip install asmd
- To install datasets:
python -m asmd.install
- To import API:
from asmd import asmd
Other examples in the paper!
- Improved inference of misalignments
- Improved the reproducibility of artificial data
- Improved the documentation
- Added ASAP group in Maestro: broken backward compatibility
- Added score for the original score
- Added midi note matching based on EITA method
- Added missing and extra notes
- General refactoring
- Added functions for set operations among datasets (union, intersections, complement and subsampling)
- Various fixes
# Skipped
- Fixed MIDI values ([0, 128) for control changes and pitches)
- Fixed metadata error while reading audio files
- Fixed pedaling for tracks that have no pedaling
- Fixed group selection
- Added get_songs
- Improved initialization of Dataset objects
- Improved documentation
- Fixed major bug in install script
- Fixed bug in conversion tool
- Removed TRIOS dataset because no longer available
- Updated ground_truth
- Improved
parallel
function - Improved documentation
- Various fixings in
get_pedaling
- Added
nframes
utility to compute the number of frames in a given time lapse - Added
group
attribute to each track to create splits in a dataset (supported in only Maestro for now) - Changed
.pyx
to.py
with cython in pure-python mode
- Added
parallel
utility to run code in parallel over a while dataset - Added
get_pianoroll
utility to get score as pianoroll - Added
sustain
,sostenuto
, andsoft
to model pedaling information - Added utilities
frame2time
andtime2frame
to ease the development - Added
get_audio_data
to get data about audio without loading it - Added
get_score_duration
to get the full duration of a score without loading it - Added another name for the API:
from asmd import asmd
- Deprecated
from asmd import audioscoredataset
- Changed the
generate_ground_truth
command line options - Easier to generate misaligned data
- Improved documentation
- Add matching of same music piece among different datasets
- Added torch.DatasetDump for preprocessing datasets and use them in pytorch
- Add new modalities (video, images)
- Add other datasets
- Refactoring of the filter function (it's a bit long now...)
[1] Simonetta, Federico ; Ntalampiras, Stavros ; Avanzini, Federico: ASMD: an automatic framework for compiling multimodal datasets with audio and scores. In: Proceedings of the 17th Sound and Music Computing Conference. Torino, 2020 arXiv:2003.01958
---
Federico Simonetta