symusic

🎉ISMIR 2024 LBD Demo Vedio: Youtube Bilibili

Symusic("Sybolic Music") is a cross-platform note level midi decoding library with lightening speed, which is hundreds of times faster (100x to 1000x depending on your file size) than mido, the main midi parsing library in python.

The library is written in cpp and based on minimidi. It offers a python binding using pybind11.

Here, we have added a tutorial.ipynb for you to learn about how to use the library.

And a document is also available. Not complete though.

Features

• You can just read a midi file like score = symusic.Score("path to midi", ttype="tick")
• Writing back to midi is now supported! score.dump_midi("path")
• Multiple time unit (ttype) is now supported (currently tick and quarter)
• The tempo attribute in the tempo event represents quarter per minute (qpm)
• We offer some batch operation functions for both Score and Track class:
  • shift_pitch(offset: int)
  • shift_velocity(offset: int)
  • shift_time(offset: float)
  • sort(key, reverse)
• You can operate each note just like you did before in python (like PrettyMidi)
• Extremely fast pickle is now supported
• .numpy() method for getting SoA data
• .filter(func, inplace) method for filtering objets in all the "List" in symusic
• A new synthesizer is now available! It comes from our another project prestosynth. Find usages in our document. Note that we will support more features of soundfont for synthesizing in the future.

Known Bugs

Strange Noise at the End of Synthesized Audio

Currently, synthesizing audio with symusic on linux gives a probability of hearing strange noise at the end of the audio. This phenomenon cannot be reproduced consistently and does not occur on windows.

You can temporarily fix this by adding a few seconds of notes at the end and removing those seconds of audio at the end.

It is important to note that the audio synthesis functionality was implemented by us rather than using an off-the-shelf project, and it is still experimental at this time. Feel free to give us feedback on any bugs you find.

Installation

Use pre-compiled version

pip install symusic

Build from source

Make sure that your system has cmake and c++ compilers

You could clone from github and install it by pip

git clone --recursive https://github.com/Yikai-Liao/symusic
pip install ./symusic

Or you could install the source distribution from pypi

pip install symusic --no-binary symusic

For debugging purpose, you could pass -Ccmake.define.MEM_LEAK_WARNING=True to pip install to enable the memory leak warning from nanobind.

pip install -Ccmake.define.MEM_LEAK_WARNING=True ./symusic

Benchmark

• midifile is writen in cpp, and could parse midi files to both event level and note level. It is slow mostly because of iostream.
• mido is writen in pure python, and only parses midi files to event level
• pretty_midi and miditoolkit is based on mido, and parse midi files to note level
• For libraries written in python or with python bindings, we use timeit to measure the time cost of parsing the midi file. nanobench for cpp libraries and BenchmarkTools for julia libraries.
• The following Common MIDI File benchmarks could be found in symusic-benchmark. And they are tested on (AMD Ryzen 7 8845H, 32GB 7500 MT/s, Linux 6.11.7)

Common MIDI File Parsing

Common MIDI File Dumping

Large MIDI File Parsing

• test using mahler.mid from minimidi/example on my laptop (i7-10875H, 32GB 2666MHz DDR4 RAM, Linux 6.1.69-1-lts)
• Note that mahler.mid is quite a large midi file. So this benchmark mainly reelects the parsing time (the percentage of time of loading file gets more significant when file is smaller)

Citation

@inproceedings{symusic2024,
    title={symusic: A swift and unified toolkit for symbolic music processing},
    author={Yikai Liao, Zhongqi Luo, et al.},
    booktitle={Extended Abstracts for the Late-Breaking Demo Session of the 25th International Society for Music Information Retrieval Conference},
    year={2024},
    url={https://ismir2024program.ismir.net/lbd_426.html#lbd},
}

Acknowledgement

• minimidi : A fast and lightweight midi parsing library written in cpp, which is the foundation of this project.
• prestosynth : A new fast soundfont synthesizer written in cpp, which is the foundation of the synthesizer in this project.
• nanobind : A efficient and lightweight library for binding C++ to Python, which is significantly faster than pybind11.
• zpp_bits : An extraordinary fast and lightweight single header library for serialization and deserialization. I use it to support pickle.
• geek_time_cpp The example code of the book "Modern C++ Programming Practice". We use the metamacro.h in it for shortening the code.
• utfcpp An easy to use and portable library for handling utf8 string in C++.