Music genre recognition using spectrograms

In this paper we present an alternative approach for music genre classification which converts the audio signal into spectrograms and then extracts features from this visual representation. The idea is that treating the time-frequency representation as a texture image we can extract features to build reliable music genre classification systems. The proposed approach also takes into account a zoning mechanism to perform local feature extraction, which has been proved to be quite efficient. On a very challenging dataset of 900 music pieces divided among 10 music genres, we have demonstrated that the classifier trained with texture compares similarly to the literature. Besides, when it was combined with other classifiers trained with short-term, low-level characteristics of the music audio signal we got an improvement of about 7 percentage points in the recognition rate.