Search

Throughout this thesis we will explore automatic music source separation by utilizing modern (at the time of writing) techniques and tools from machine learning and big data processing. The bulk of this work was carried out between 2016 and 2019. In Chapter 2 we conduct a review of source separation literature. We start by outlining a subset of applications of source separation in some depth. We describe some of the early, pioneering work in automatic source separation: Auditory Scene Analysis, and its digital counterpart, Computational Auditory Scene Analysis. We then introduce matrix decomposition-based methods such as Independent Component Analysis and Non-Negative Matrix factorization, and pitch informed methods where the separation algorithm is guided by pitch information that is known a priori. We brie y discuss user-guided methods, before conducting a thorough review of Deep Learning based source separation, including recurrent, convolutional, deep clustering-based, and Generative Adversarial Networks. We then proceed to describe common evaluation metrics and training datasets. Finally, we list a number of current challenges and drawbacks of current systems. Chapter 3 focuses on datasets for musical source separation. First we show the growth of dataset sizes for both machine learning in general and music information retrieval specifically. We give several examples of the complexities and idiosyncrasies that are intrinsic to music datasets. We then proceed to present a method for extracting ground truth data for source separation from large unstructured musical catalogs. In Chapter 4 we design a novel deep learning-based source separation algorithm. Motivation is provided by means of a musicological study1 that showed the high importance of vocals relative to other musical factors, in the minds of listeners. At the core of the vocal separation algorithm is the U-Net, a deep learning architecture that uses skip connections to preserve fine-grained detail. It was originally developed in the biomedical imaging domain, and later adapted to image-to-image translation. We adapt it to the source separation domain by treating spectrograms as images, and we use the dataset mining methods from Chapter 3 to generate sufficiently large training data. We evaluate our model objectively using standard evaluation metrics, subjectively using \crowdsourced" human subjects. To the best of our knowledge, this is the first use of U-Nets for source separation. In the introduction above we proposed joint learning to optimize source separation and other objectives. In Chapter 5 we investigate one such instance: multi-task learning of vocal removal and vocal pitch tracking. We combine the vocal separation model from Chapter 4 with a state of the art pitch salience estimation model2, exploring several ways of combining the two models. We find that vocal pitch estimation benefits from joint learning when the two tasks are trained in sequence, with the source separation model preceding the pitch estimation model. We also report benefits from fine-tuning by iteratively applying the model. Chapter 6 extends the U-Net model to multiple instruments. In order to minimize the phase artifacts that were a common issue in Chapter 4, we modify the model to operate in the complex domain. We run experiments with several loss functions: Time-domain loss, magnitude-only frequency domain loss, and joint time and frequency-domain loss. Our experiments are evaluated both objectively and subjectively, and we carry out extensive qualitative analysis to investigate the effects of complex masking. Finally, we conclude the thesis in Chapter 7 by summarizing this work and highlighting several future directions of research.