Automatic musical genre recognition using deep convolutional neural networks

Abstract

For the long time in computer vision and digital signal processing manually developed algorithms and filters were used. With the development of computers technics and constantly growing amount of available data samples, these algorithms became less accurate than modern machine learning approaches. The idea behind them is to construct useful representations based on data itself rather than on expert knowledge. Such approach allows machine learning algorithms to choose for themselves which parts of data more important. Today machine learning is successfully applied in such tasks as image recognition in Google image search and speech recognition in Google Now, Siri, Cortana. Nowadays best approaches are built upon different variations of neural network algorithms. One of the fields, where machine learning are successfully applied is music information retrieval, where musical genres classification is one of the main tasks and solving it efficiently can help automatically organize large collections of musical data which are available for now. As music genre aggregates a lot of song information, model for calculating music song similarities based on audio information can possibly be built on proposed model. In this article, the algorithms for automatic music genre recognition are discussed and usage of deep convolutional neural networks is proposed for this task. The network’s architecture is described and its quality evaluated on real-world data. In this work GTZAN dataset is used and classification problem for four and ten genres classification was examined using mel-frequency cepstral coefficients and waveform as features. The quality of proposed algorithm was evaluated on hold-out set for four and ten different genres and compared to using restricted Boltzmann machines for four genres classification. The resulting accuracy for our genres classification task is 76%, which is about 15% better than restricted Boltzmann machine approach. Though model overfits strongly on rather small dataset it can be fixed by using larger amount of data. The main differences between proposed neural network architecture and traditional convolutional neural networks are gated activations, dilated convolutions and residual connections. Gated activations allow the network to additionally weight and inhibit importance of intermediate features like it is done in recurrent neural networks. Dilated convolutions allow increasing receptive field of network’s filters while maintaining small number of trainable parameters. Residual connections are proven to be vital feature for very deep neural networks to prevent gradient degrading and neural networks with residual connections yields best classifications accuracy for image classifications task for now. The proposed neural network is used to classify musical genres, based on pure waveform or mel-frequency cepstral coefficients, which are well known to be good sound representation for speech recognition task. Ref. 11, fig. 3, tabl. 2.

Keywords

References

1. H. Lee, P. Pham, Y. Largman, and A. Y. Ng, “Unsupervised feature learning for audio classification using convolutional deep belief networks,” in Advances in Neural Information Processing Systems 22, Curran Associates, Inc., 2009, pp. 1096–1104, URL: http://papers.nips.cc/paper/3674-unsupervised-feature-learning-for-audio-classification-using-convolutional-deep-belief-networks.pdf.
2. T. Feng, “Deep learning for music genre classification,” URL: https://courses.engr.illinois.edu/ece544na/fa2014/Tao_Feng.pdf.
3. M. Haggblade, Y. Hong, and K. Kao, “Music Genre Classification,” Stanfor, 2009, URL: https://cs229.stanford.edu/Fproj2011/HaggbladeHongKao-MusicGenreClassification.pdf&usg=AOvVaw1Nc3dcWm4P4rW6UsiLu712.
4. G. E. Hinton, S. Osindero, and Y.-W. Teh, “A Fast Learning Algorithm for Deep Belief Nets,” Neural Comput., vol. 18, no. 7, pp. 1527–1554, 2006, DOI: 10.1162/neco.2006.18.7.1527.
5. Y. Cai, D. Ji, and D. Cai, “A KNN Research Paper Classification Method Based on Shared Nearest Neighbor,” in NTCIR-8 Workshop Meeting, 2010, pp. 336–340.
6. T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, and A. Y. Wu, “An efficient k-means clustering algorithm: analysis and implementation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 881–892, 2002, DOI: 10.1109/TPAMI.2002.1017616.
7. M. Mandel and D. Ellis, “Song-Level Features And Support Vector Machines For Music Classification,” in 6th International Conference on Music Information Retrieval, 2005, pp. 594–599, URL: http://ismir2005.ismir.net/proceedings/1106.pdf.
8. I. Goodfellow, Y. Bengio, and A. Courville, “Deep feedforward networks,” in Deep learning, The MIT Press, 2016, pp. 164–223, URL: http://worldcat.org/isbn/9780262035613.
9. G. Tzanetakis, G. Essl, and P. Cook, “Automatic Musical Genre Classification Of Audio Signals,” in 2nd Annual International Symposium on Music Information Retrieval 2001, 2001, URL: http://ismir2001.ismir.net/pdf/tzanetakis.pdf.
10. J. Leben, “Data Sets,” MARSYAS, 2001. [Online]. Available: http://marsyas.info/downloads/datasets.html.
11. A. van den Oord et al., “WaveNet: A Generative Model for Raw Audio,” Sep. 2016, arXiv: 1609.03499v2.

License

Copyright (c) 2018 Ярослав Юрійович Дорогий, Василь Васильович Цуркан, Олександр С. Хапілін (Автор)