Variational blind audio deconvolution with a Gaussian process kernel bank

Presentation at the fantastic BIASLab, TU Eindhoven. Abstract below; find the slides here.

Blind audio deconvolution is the problem of recovering an unknown source and filter from their convolution alone, given only the output audio signal. As a textbook example of a severely underdetermined problem, it is in heavy need of some proper regularization.

We attack it via Bayesian regularization on both sides. On the source side, we learn a data-driven prior directly from a large corpus in the form of a Gaussian process filterbank. On the filter side, from a more theoretical angle, we introduce a new informative prior for AR($p$) coefficients.

We then apply this to speech, where blind audio deconvolution is known as glottal inverse filtering, with main applications in clinical medicine. Our variational solution admits principled inference in realtime on a single GPU, and reaches state of the art on the benchmarks we tested.

Go to next post ▤ in research series.