Log-Mel Spektrogramlar Kullanılarak Müzik Türü Sınıflandırmasında CNN, CRNN ve Residual CNN Modellerinin Performans Karşılaştırması

Öz

Dijital müzik arşivlerinin hızla büyümesi, müzik içeriklerinin otomatik olarak sınıflandırılması ve düzenlenmesi ihtiyacını artırmıştır. Bu çalışmada, müzik türü sınıflandırması amacıyla Evrişimsel Sinir Ağı (CNN), Evrişimsel Tekrarlayan Sinir Ağı (CRNN) ve Artık Bağlantılı Evrişimsel Sinir Ağı (Residual CNN) modellerinin performansları karşılaştırılmıştır. Deneysel çalışmalarda blues, klasik, country, disco, hip-hop, caz, metal, pop, reggae ve rock olmak üzere 10 müzik türünden oluşan GTZAN veri seti kullanılmıştır. Veri seti, her biri yaklaşık 30 saniye uzunluğunda toplam 1000 ses kaydını içermektedir. Ses kayıtları 22.050 Hz örnekleme frekansında tek kanallı olarak işlenmiş ve her kayıt örtüşmesiz 3 saniyelik alt parçalara ayrılmıştır. Böylece toplam 10.000 ses segmenti elde edilmiştir. Veri sızıntısını önlemek amacıyla aynı müzik kaydından elde edilen bütün segmentler yalnızca eğitim, doğrulama veya test kümelerinden birinde tutulmuştur. Ses sinyalleri, 2048 örnek FFT pencere uzunluğu, 512 örnek ilerleme miktarı ve 128 Mel filtresi kullanılarak Log-Mel spektrogramlara dönüştürülmüştür. CNN, CRNN ve Residual CNN modelleri aynı veri bölümleri, ön işleme adımları ve eğitim parametreleri kullanılarak eğitilmiştir. Modellerin performansları doğruluk, kesinlik, duyarlılık, makro F1 skoru ve ağırlıklı F1 skoru ölçütleriyle değerlendirilmiştir.

Anahtar Kelimeler

• Müzik türü sınıflandırması
• derin öğrenme
• CNN
• CRNN
• Residual CNN
• Log-Mel spektrogram

Performance Comparison of CNN, CRNN and Residual CNN Models in Music Genre Classification Using Log-Mel Spectrograms

Öz

The rapid growth of digital music archives has increased the need for automatic classification and organization of music content. In this study, the performances of Convolutional Neural Network, Convolutional Recurrent Neural Network, and Residual Convolutional Neural Network models were compared for music genre classification. The GTZAN dataset, consisting of ten music genres including blues, classical, country, disco, hip-hop, jazz, metal, pop, reggae, and rock, was used in the experiments. The dataset contains a total of 1,000 audio recordings, each approximately 30 seconds long. The audio recordings were processed in mono format at a sampling frequency of 22,050 Hz, and each recording was divided into non-overlapping three-second segments. Thus, a total of 10,000 audio segments were obtained. To prevent data leakage, all segments derived from the same original recording were included in only one of the training, validation, or test sets. The audio signals were transformed into Log-Mel spectrograms using an FFT window length of 2,048 samples, a hop length of 512 samples, and 128 Mel filters. The CNN, CRNN, and Residual CNN models were trained using identical data partitions, preprocessing procedures, and training parameters. Model performances were evaluated using accuracy, precision, recall, macro F1-score, and weighted F1-score.

Anahtar Kelimeler

• Music genre classification
• deep learning
• CNN
• CRNN
• Residual CNN
• Log-Mel spectrogram

Kaynakça

1. Ashraf, M., Abid, F., Raza, M. O., Rasheed, J., Alsubai, S., ve Asuroglu, T. (2025). Music genre classification with modified residual learning and dual neural network. PLOS ONE, 20(10 October). https://doi.org/10.1371/journal.pone.0333808
2. Chathuranga, D., ve Jayaratne, L. (2013). Automatic Music Genre Classification of Audio Signals with Machine Learning Approaches. GSTF Journal on Computing (JoC), 3(2), 14. https://doi.org/10.7603/s40601-013-0014-0
3. Choi, K., Fazekas, G., Sandler, M., ve Cho, K. (2017). Convolutional recurrent neural networks for music classification. 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2392–2396. https://doi.org/10.1109/ICASSP.2017.7952585
4. Dieleman, S., ve Schrauwen, B. (2014). End-to-end learning for music audio. 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 6964–6968. https://doi.org/10.1109/ICASSP.2014.6854950
5. Ding, Y., Zhang, H., Huang, W., Zhou, X., ve Shi, Z. (2024). Efficient Music Genre Recognition Using ECAS-CNN: A Novel Channel-Aware Neural Network Architecture. Sensors, 24(21). https://doi.org/10.3390/s24217021
6. Li, T. (2024). Optimizing the configuration of deep learning models for music genre classification. Heliyon, 10(2), e24892. https://doi.org/10.1016/J.HELIYON.2024.E24892
7. Srivastava, N., Ruhil, S., ve Kaushal, G. (2022). Music Genre Classification using Convolutional Recurrent Neural Networks. 2022 IEEE 6th Conference on Information and Communication Technology (CICT), 1–5. https://doi.org/10.1109/CICT56698.2022.9997961
8. Sturm, B. L. (2012). An analysis of the GTZAN music genre dataset. Proceedings of the Second International ACM Workshop on Music Information Retrieval with User-Centered and Multimodal Strategies, MIRUM ’12, 7–12. https://doi.org/10.1145/2390848.2390851

9. Tzanetakis, G., ve Cook, P. (2002). Musical genre classification of audio signals. IEEE Transactions on Speech and Audio Processing, 10(5), 293–302. https://doi.org/10.1109/TSA.2002.800560
10. Wang, Z., Muknahallipatna, S. S., Fan, M., Okray, A., ve Lan, C. (2019). Music Classification using an Improved CRNN with Multi-Directional Spatial Dependencies in Both Time and Frequency Dimensions. 2019 International Joint Conference on Neural Networks (IJCNN), 1–8. https://api.semanticscholar.org/CorpusID:203605838