In the last two years, new developments in acoustic modeling with artificial neural networks (ANN) have led to increases in performance against conventional Gaussian mixture model architectures of around 20%-30% relative. These improvements, initially obtained by Microsoft research, have now been reproduced and confirmed by Google. This last company is, in fact, already using these new models in their speech recognition system. Given the fact that our in house speech recognition system uses classic ANN technology for acoustic modeling, these new developments are worth exploring. The talk will introduce the deep belief network implementation that has led to the aforementioned improvements and compare it to our MLP implementation. The use of GPUs will also be brought into discussion since this seems to be a key point for development. This talk will not be high profile but rather an starting point for an exchange of ideas/expertise between those interested in using/developing AUDIMUS.
Note: This seminar will be held in English, if required.