Generative Model of Voice in Noise for Structured Coding Applications

A generative model of a human voice is presented, based on many pseudo-physical considerations. For robustness, observation noise is also included in the model. An EM-algorithm framework for inference and learning is then described. An instance of approximate inference and subsequent learning presented allows an extraction of voice parameter which can be used for structured coding application. This set of parameters allows a great amount of compression as well as the flexibility in making modification to pitch, duration and breathiness, noise-free synthesis compared to other non-parametric approaches.