Turkish music genres classification using convolutional neural network

Music genre classification, frequently used to organize digital music databases, has become a challenging task in the field of music information retrieval due to the exponential growth of digital music recordings. Genre classification can be useful in addressing certain current interesting challenges, such as developing song recommendations, finding similar songs, and discovering audiences who might enjoy that specific music. Therefore, automatic music genre classification plays an important role for applications that offer music analysis, retrieval, and recommendation systems. It can assist or even replace human users in this process. Furthermore, automatic musical genre classification serves as a framework for developing and assessing features for any type of content-based analysis of musical signals. Previous research in this field has primarily focused on exploring the classification of genres in Western music. However, categorizing online Turkish music content remains inadequately defined, posing a challenge for the automatic classification of audio genres in the Turkish music context. The earlier work on Turkish music genre classification relies mostly on classical techniques for genre classification. These classical techniques depend on two basic steps: the first is extracting the acoustic features of music genre and the second is using machine learning algorithms for classification, which limits the performance of Turkish music genre classification. Therefore, in this research, automatic Turkish music genre classification with convolutional neural networks by using images obtained by the Mel-spectrogram analysis is studied. The initial step involves the creation of a dataset comprising of the most six well-known categories of Turkish music, which include: Arabesque, Pop, Rap, Rock, Turkish Classical Music, and Turkish Religious Music. Then, this research proposes a deep learning technique with convolutional neural networks and employs spectrogram images produced from Mel-Spectrogram as the input into a CNN to classify the songs into the appropriate musical genres. The proposed method has an accuracy level of about 99.62% for training and 99.80% for testing, which confirms the success of this method.