Automatic Music Genre Classification

1 – Introduction In this work, we are presenting our approach to automatic genre classification for music files, or songs, which consists of audio files represented by a time series data, where the goal is to automatically process the files, to establish a genre assignment. Such applications that require automatic genre classification include internet radio stations that play similar songs based on a user specified preference, or for automatic organization of music databases, where large number of songs must be organized and categorized based on style, in addition to the artist information. This task of genre classification involves processing time series data, but also most significantly involves the classification of subjective data, where some songs might fall into multiple categories, or some categories might be similar to subjective observers. This categorization issue can be explained by a taxonomical analysis of the categories and the multiple genres assigned to music, where for instance some songs might be considered a particular style just based on the artist, even though the particular song diverges from the typical assignment. Our contribution to the audio processing community, involves an implementation of a successful genre classification algorithm, that as we will describe in detail, consists of a feature extraction process that utilizes Mel-Frequency Cepstral Coefficients (MFCCs) and Spectral and cepstral parameters, and then the data is processed by a Linear Discriminant analysis technique to reduce the dimensionality of our feature vector, such that finally we can then utilize a machine learning technique that consists of a quadratic classifier as an eager learner that processes training data that is pre-labeled by a user trainer, such that the classifier parameters can be learned and estimated for establishing the potential to generalize and classify novel and new unseen test data, or any music file, where based on our testing results, as we will describe in section 5, we are confident that our process will perform adequately for general test cases.