A Multiclass Semi-Supervised Deep Convolutional Generative Adversarial Network for Music Genre Classification Using Mel-Frequency Cepstral Coefficients

The growing consumer base and expanding market for various music styles highlight the necessity of classifying music genres to cater to people's preferences. Manual music ranking is a labor-intensive process for listeners, prompting the need for a more efficient approach. This involves the extr...

Full description

Saved in:
Bibliographic Details
Published in2024 International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics (IITCEE) pp. 1 - 6
Main Authors Satapathy, Ashutosh, Duda, Neelima, Machcha, Jaswanthi, Thottempudi, Kokila
Format Conference Proceeding
LanguageEnglish
Published IEEE 24.01.2024
Subjects
Computational modeling
Data augmentation
Feature extraction
Generative adversarial networks
GTZAN
Indian Music Genre
Mel-frequency Cepstral Coefficients
Multiclass Classification
Music
Semi-Supervised Deep Convolutional Generative Adversarial Network
Training
Training data
Online AccessGet full text

Cover

More Information
Summary:The growing consumer base and expanding market for various music styles highlight the necessity of classifying music genres to cater to people's preferences. Manual music ranking is a labor-intensive process for listeners, prompting the need for a more efficient approach. This involves the extraction of Mel-frequency cepstral coefficient feature maps from the log Mel-spectrograms of audio clips. The extracted feature maps are supplied to a multiclass semi-supervised deep convolutional generative adversarial network where the discriminator behaves as a classifier. The training of models involves utilizing the GTZAN standardized dataset, a publicly accessible collection of thousands of audio files spanning ten different genres, from which 80% and 20% of the data are used for training and testing, respectively. Finally, the paper discusses the performance of the semi-supervised deep convolutional generative adversarial network through the RMSprop and Adam optimizers on the original and augmented labeled and unlabeled MFCC feature maps. Without any data augmentation, the discriminator achieves a training accuracy of 97.9% and a test accuracy of about 45.67%. In contrast, the discriminator's training accuracy is about 98.3%, and the test accuracy is 84.75% with data augmentation.
DOI:10.1109/IITCEE59897.2024.10467652