This thesis is concerned with methodological advances in probabilistic inference and their application to core challenges in machine perception and AI. Inferring a posterior distribution over the parameters of a model given some data is a central challenge that occurs in many fields ranging from finance and artificial intelligence to physics. Exact calculation is impossible in all but the simplest cases and a rich field of approximate inference has been developed to tackle this challenge. This thesis develops both an advance in approximate inference and an application of these methods to the problem of speech synthesis. In the first section of this thesis we develop a novel framework for constructing Markov Chain Monte Carlo (MCMC) kernels that can efficiently sample from high dimensional distributions such as the posteriors, that frequently occur in machine perception. We provide a specific instance of this framework and demonstrate that it can match or exceed the performance of Hamiltonian Monte Carlo without requiring gradients of the target distribution. In the second section of the thesis we focus on the application of approximate inference techniques to the task of synthesising human speech from text. By using advances in neural variational inference we are able to construct a state of the art speech synthesis system in which it is possible to control aspects of prosody such as emotional expression from significantly less supervised data than previously existing state of the art methods.