Efficient Music Genre Classification for Real-Time Recommender Systems

All across the world, different people tend to listen to different music genres or subsets of genres that are similar to one another, according to their mood, the weather, or the occasion. At different times, people like to listen to distinct playlists, depending on circumstances such as the ones mentioned above. The main idea underlying the work in this thesis is to build and test a variety of ideas for automatically classifying music samples according to their genres, also known as Automatic Genre Recognition. Here, the main aim is to minimize the time for training the model and feature generation, as it is one of the main challenges in the field of Musical Information Retrieval systems. There have been numerous music recommendation systems based on music tagging by experts, and many other systems analyze other users’ behavior in order to generate recommendations. However, it would be useful to create a music recommendation system purely based on musical information retrieval. The algorithm devised in this thesis can be used for fast training while providing good classification performance. This is achieved by methods such as reducing the number of features required to train the model and minimizing the length of musical clips to be processed for generating those features. By classification and identifying the particular user’s behavior, this algorithm can generate a smart play list in shuffle mode that is likely to include songs that the user might like to listen to under the given circumstances