Back to Search Start Over

NLP-based music processing for composer classification.

Authors :
Deepaisarn, Somrudee
Chokphantavee, Sirawit
Chokphantavee, Sorawit
Prathipasen, Phuriphan
Buaruk, Suphachok
Sornlertlamvanich, Virach
Source :
Scientific Reports; 8/14/2023, Vol. 13 Issue 1, p1-10, 10p
Publication Year :
2023

Abstract

Categorizing music pieces by composer is a challenging task in digital music processing due to their highly flexible structures, introducing subjective interpretation by individuals. This research utilized musical data from the MIDI and audio edited for synchronous tracks and organization dataset of virtuosic piano pieces. In this work, pitch and duration were the musical features of interest. The goal was to innovate an approach to representing a musical piece using SentencePiece and Word2vec, which are natural language processing-based techniques. We attempted to find correlated melodies that are likely to be formed by single interpretable units of music via co-occurring notes, and represented them as a musical word/subword vector. Composer classification was performed in order to ensure the efficiency of this musical data representation scheme. Among classification machine learning algorithms, k-nearest neighbors, random forest classifier, logistic regression, support vector machines, and multilayer perceptron were employed to compare performances. In the experiment, the feature extraction methods, classification algorithms, and music window sizes were varied. The results were that classification performance was sensitive to feature extraction methods. Musical word/subword vector standard deviation was the most effective feature, resulting in classification with a high F1-score, attaining 1.00. No significant difference was observed among model classification performances. [ABSTRACT FROM AUTHOR]

Details

Language :
English
ISSN :
20452322
Volume :
13
Issue :
1
Database :
Complementary Index
Journal :
Scientific Reports
Publication Type :
Academic Journal
Accession number :
169943890
Full Text :
https://doi.org/10.1038/s41598-023-40332-0