Automatic music recommendation requires music information retrieval tasks ranging from classifying genre and mood, to artist relatedness. In this thesis, we propose using deep learning models for music classification tasks and provide both the motivation and analysis of why deep learning models are suitable for these tasks.

Similar to many tasks that require supervised learning, music information retrieval required an elaborate feature engineering procedure that depends heavily on human labor. Many previous research has been focused on discovering the most "salient" features for music. Such quest has been proven to be challenging. In this thesis, we investigate neural networks in order to explore its ability to automatically engineer features without human intervention. Con...[ Read more ]

Automatic music recommendation requires music information retrieval tasks ranging from classifying genre and mood, to artist relatedness. In this thesis, we propose using deep learning models for music classification tasks and provide both the motivation and analysis of why deep learning models are suitable for these tasks.

Similar to many tasks that require supervised learning, music information retrieval required an elaborate feature engineering procedure that depends heavily on human labor. Many previous research has been focused on discovering the most "salient" features for music. Such quest has been proven to be challenging. In this thesis, we investigate neural networks in order to explore its ability to automatically engineer features without human intervention. Convolutional Neural Networks have been shown to classify images directly from pixels in the image recognition area. It has been shown that "neurons' in an CNN can be driven by the raw input data adapt the final model to a desirable state, in a supervised learning paradigm.

In this thesis, we show for the first time that deep learning models can learn directly from raw time-domain audio data and bypass both signal processing and feature engineering steps in music classification tasks. We carry out our experiments both on audio clips and lyrics. We compare the results from models learned from raw data to those from feature-engineered data, showing slightly better results with raw audio input. Moreover, we give a full analysis on the output of each convolution layer showing that the CNN is indeed performing signal processing and feature extractor automatically. We also show that CNN with word embeddings, without lexicon features, can be used to directly classify lyrics from words. Our results on using CNN to classify music from raw time-domain audio data is later on applied to a task of speech emotion recognition, showing its generality.