Personal tools
You are here: Home Events ANC Workshop: Benigno Uria and Iain Murray, Chair: Matthias Hennig

ANC Workshop: Benigno Uria and Iain Murray, Chair: Matthias Hennig

— filed under:

  • ANC Workshop Talk
When Jan 15, 2013
from 11:00 AM to 12:00 PM
Where IF 4.31/4.33
Add event to calendar vCal

Benigno Uria

Connectionist probabilistic models for speech processing

Deep neural networks have obtained big accuracy improvements in speech recognition during the last 5 years. Most of these deep methods use a probabilistic model to obtain features that are then fed to a classifier. However, a better probabilistic model does not guarantee a better set of classification features.
In my present work I'm trying to use connectionist probabilistic models directly as emission distributions of hidden Markov models. Hopefully, this will make a better probabilistic model translate directly into better speech recognition accuracies.


Iain Murray

Revisiting multivariate density estimation

I'm currently working on a project to measure some properties of the Milky Way. As often happens, part of the modelling process requires fitting a multivariate probability distribution. I found the text-book baselines of mixtures of Gaussians (or factor analyzers) to be inadequate. A (fairly) simple autoregressive technique seems to be working well. I will recommend considering this style of approach (including a more advanced version that Benigno UrĂ­a is working on) for use in a range of data-analysis problems. That is, those interested in spike trains, images, general machine learning, etc. may find this stuff useful.