A MIDI and music data manipulation library
Project description
S-Coda
Overview
S-Coda is a Python library for handling MIDI files. It is written with the purpose of machine learning tasks in mind. S-Coda supports a plethora of different MIDI manipulation operations, such as:
- quantisation of notes
- quantisation of note lengths
- splitting sequences into bars
- transposing of sequences
- creating piano-roll visualisations of pieces
- tokenisation of sequences
S-Coda was used in our project PAUL-2 to process MIDI files. For information about how to use S-Coda we refer to chapter 5 of the thesis in which S-Coda was introduced. Note that this thesis refers to version 1.0 of S-Coda, which has since received major overhauls.
Installation
We recommend installing S-Coda from PyPI using pip:
pip install scoda
Changelog
See CHANGELOG.md for a detailed changelog.
Usage
Example
We refer to the aforementioned thesis for a more in-depth guide on how to use S-Coda. We provide a short listing on how to use basic S-Coda functions, which is up-to-date as of version 2.4.0.:
# Load sequence, choose correct track (often first track contains only meta messages)
sequence = Sequence.sequences_load(file_path=RESOURCE_BEETHOVEN)[1]
# Quantise the sequence to sixteenths and sixteenth triplets (standard values)
sequence.quantise_and_normalise()
# Split the sequence into bars based on the occurring time signatures
bars = Sequence.sequences_split_bars([sequence], meta_track_index=0)[0]
# Prepare tokeniser and output tokens
tokeniser = MultiTrackLargeVocabularyNotelikeTokeniser(num_tracks=1)
tokens = []
# Tokenise all bars in the sequence
for bar in bars:
tokens.extend(tokeniser.tokenise([bar.sequence]))
# Convert to a numeric representation
encoded_tokens = tokeniser.encode(tokens)
# (Conduct ML operations on tokens)
encoded_tokens = encoded_tokens
# Convert back to token representation
decoded_tokens = tokeniser.decode(encoded_tokens)
# Create sequence from tokens
detokenised_sequences = tokeniser.detokenise(decoded_tokens)
# Save sequence
detokenised_sequences[0].save("../out/generated_sequence.mid")
Implementational Details
S-Coda is built around the Sequence class, which represents a musical sequence.
The Sequence object is a wrapper for two internal classes, AbsoluteSequence and RelativeSequence, which represent
music in two different ways.
For the absolute sequences, the elements of the sequences are annotated with their absolute points in time within the
sequence, while for the relative sequence elements specify the time between events.
These two representations are used internally for different operations.
The Sequence object abstracts away the differences between these two representations and provides the user with a
unified experience.
The atomic element of S-Coda is the Message, which is comparable to a MIDI event.
Messages have a MessageType, denoting the type of the message, and several other fields depending on which type of
message it is.
For example, a message of type NOTE_ON will have a note field, which denotes the pitch number of the note that it
represents.
Note that directly editing single messages or the messages of a sequence is possible, but not recommended, as it can
lead to inconsistencies in the Sequence object.
If you still need to do so, make sure to invalidate either the absolute or relative internal representation (using
Sequence.invalidate_abs() and Sequence.invalidate_rel()) after directly editing messages.
This is not required when modifying the sequence using the functions provided by Sequence, as staleness of the
internal representations is kept track of this way.
Citing
If you use S-Coda in your research, please cite the following paper:
@inproceedings{Schoen.2023,
author = {Felix Sch{\"{o}}n and
Hans Tompits},
title = {{PAUL-2:} An Upgraded Transformer-Based Redesign of the Algorithmic Composer {PAUL}},
booktitle = {22nd International Conference of the Italian Association for Artificial Intelligence ({AIxIA 2023})},
series = {Lecture Notes in Computer Science},
volume = {14318},
pages = {278--291},
publisher = {Springer},
year = {2023},
doi = {10.1007/978-3-031-47546-7\_19}
}
