MATH 488/489: Stephen Marsland's Projects
Generative models of birdsong
Generative models produce samples that look like the data they were trained on. For sounds, this means that they should produce things that sound correct. We’ll explore the theory of this a little, and then see if it is any good in practice.
Given a set of time series data that are the same qualitatively, but have some time-variation, one way to line them up is to reparameterise one of them. We will investigate a method of doing this called the square-root velocity transform, which is the application of a diffeomorphism of the line to the data.
