Automatic Speech Processing by Inference in Generative Models

Summary

In this chapter, we have explored the use of inference in probabilistic generative models as a powerful signal processing tool for speech and audio. The basic paradigm explored was to design a simple model for the data we observe in which the key quantities that we would eventually like to compute appear as hidden (latent) variables. By executing probabilistic inference in such models, we automatically estimating the hidden quantities and thus perform our desired computation. In a sense, the rules of probability derive for us, automatically, the optimal signal processing algorithm for our desired outputs given our inputs under the model assumptions. Crucially, even though the generative model may be quite simple and may not capture all of the variability present in the data, the results of inference can still be extremely informative.

We gave several examples showing how inference in very simple generative models can be used to perform surprisingly complex speech processing tasks including denoising, source separation, pitch tracking, timescale modification and estimation of articulatory movements from audio.