Fundamental frequency estimation is the process of estimating time-varying fundamental frequency values from musical audio, and encompasses problems including pitch tracking, melody extraction, bass line estimation, and multiple-f0 estimation. Fundamental frequency estimation is challenging for a number of reasons. First, it aims to measure a perceptual quantity from audio that is the result of a complex series of processing stages in the brain. Second, it involves performing estimations in the presence interfering sources, some of which may be highly correlated with the target source. Until recently, there has been relatively little work on data-driven f0 estimation, primarily due to the lack of annotated f0 data. Furthermore, each type of f0 estimation has been researched as an isolated research area, with the majority of algorithms designed for and tested on a single musical target, even if the approach could be easily extended to other contexts. In this dissertation we address the following questions: (1) How can we obtain large amounts of annotated f0 data to train and evaluate data-driven methods? (2) Are data-driven methods more accurate than the traditional approaches to f 0 estimation? (3) How can we exploit the inherent relationships between different types of f0 estimation? and (4) How can we use f0 information to address higher level tasks such as singing style analysis? A primary contribution of this dissertation is the development of the MedleyDB multitrack dataset with f 0 annotations, and methodologies for leveraging and augmenting the data to be used for f0 estimation. The second main contribution is a set of extensive experiments on data-driven models for f0 estimation including convolutional neural networks and multitask learning that lead to an improvement on the state of the art for a number of problems in f0 estimation.