Implementasi Klasifikasi Genre Musik Pada Dataset GTZAN Menggunakan Convolutional Neural Network Berbasis Mel-spectrogram Dengan Augmentasi Data

Rizqullah, Fadhiil Hanif (2026) Implementasi Klasifikasi Genre Musik Pada Dataset GTZAN Menggunakan Convolutional Neural Network Berbasis Mel-spectrogram Dengan Augmentasi Data. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221144-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Perkembangan platform musik digital menyebabkan kebutuhan akan sistem klasifikasi genre musik otomatis yang mampu mengelola koleksi musik dalam jumlah besar secara efektif. Penelitian ini bertujuan untuk menganalisis penerapan Mel-spectrogram sebagai representasi fitur audio, membandingkan kinerja model Convolutional Neural Network (CNN), ResNet50, dan VGG16, serta mengevaluasi pengaruh augmentasi data terhadap performa klasifikasi genre musik. Dataset yang digunakan adalah GTZAN Music Genre Dataset yang terdiri atas 1.000 audio dari 10 genre musik. Tahap preprocessing meliputi resampling 22.050 Hz, normalisasi amplitudo, segmentasi audio 10 detik, dan ekstraksi Mel-spectrogram dengan 128 pita Mel. Penelitian menggunakan delapan skenario pengujian, yang mencakup satu skenario tanpa augmentasi dan tujuh skenario dengan penerapan augmentasi pitch shifting, noise injection, time stretching, maupun kombinasi dari ketiganya. Hasil penelitian menunjukkan bahwa Mel-spectrogram mampu merepresentasikan karakteristik genre musik secara efektif, ditunjukkan oleh accuracy di atas 86% pada seluruh model tanpa augmentasi. Pengaruh augmentasi berbeda pada setiap model. CNN dan ResNet50 memperoleh performa terbaik pada kombinasi pitch shifting dan time stretching dengan accuracy masing-masing sebesar 91,54% dan 90,20%. Sementara itu, VGG16 menghasilkan performa terbaik pada kombinasi pitch shifting, noise injection, dan time stretching dengan accuracy sebesar 93,32%, precision sebesar 93,40%, recall sebesar 93,33%, dan F1-score sebesar 93,16%. Hasil tersebut menunjukkan bahwa pendekatan transfer learning menggunakan VGG16 merupakan metode yang paling efektif untuk klasifikasi genre musik berbasis Mel-spectrogram pada dataset GTZAN.
=====================================================================================================================================
The rapid growth of digital music platforms has increased the need for automatic music genre classification systems capable of managing large music collections efficiently. This study aims to analyze the application of Mel-spectrograms as audio feature representations, compare the performance of Convolutional Neural Network (CNN), ResNet50, and VGG16 models, and evaluate the impact of data augmentation on music genre classification performance. The dataset used in this study was the GTZAN Music Genre Dataset, which consists of 1,000 audio tracks from 10 music genres. The preprocessing stage included resampling to 22,050 Hz, amplitude normalization, 10-second audio segmentation, and Mel-spectrogram extraction using 128 Mel bands. The study employed eight experimental scenarios, consisting of one baseline scenario without augmentation and seven scenarios applying pitch shifting, noise injection, time stretching, and their combinations. The results showed that Mel-spectrograms effectively represented music genre characteristics, as indicated by classification accuracies above 86% for all models without augmentation. The impact of data augmentation varied across models. CNN and ResNet50 achieved their best performance using the combination of pitch shifting and time stretching, with accuracies of 91.54% and 90.20%, respectively. Meanwhile, VGG16 achieved the best overall performance using the combination of pitch shifting, noise injection, and time stretching, with an accuracy of 93.32%, precision of 93.40%, recall of 93.33%, and F1-score of 93.16%. These findings indicate that the transfer learning approach based on VGG16 is the most effective method for Mel-spectrogram-based music genre classification on the GTZAN dataset.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep Learning, GTZAN, Mel-spectrogram, Musik, Deep Learning, GTZAN, Mel-spectrogram, Music
Subjects:	M Music and Books on Music > M Music T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.546 Computer algorithms
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Fadhiil Hanif Rizqullah
Date Deposited:	25 Jul 2026 04:48
Last Modified:	25 Jul 2026 04:48
URI:	http://repository.its.ac.id/id/eprint/137279

Actions (login required)

View Item