Kris Demuynck

Missing Data Solutions for Robust Speech Recognition

Current automatic speech recognisers rely for a great deal on statistical models learned from training data. When they are deployed in conditions that differ from those observed in the training data, the generative models are unable to explain the incoming data and poor accuracy results. A very noticeable effect is deterioration due to background noise. In the MIDAS project, the state-of-the-art in noise robustness was advanced on two fronts, both making use of the missing data approach. First, novel sparse exemplar-based representations of speech were proposed. Compressed sensing techniques were used to impute noise-corrupted data from exemplars. Second, a missing data approach was adopted in the context of a large vocabulary speech recogniser, resulting in increased robustness at high noise levels without compromising on accuracy at low noise levels. The performance of the missing data recogniser was compared with that of the Nuance VOCON-3200 recogniser in a variety of noise conditions observed in field data.

Designing syllable models for an HMM based speech recognition system

In automatic speech recognition, as in many areas of machine learning, stochastic modeling relies on neural networks more and more. Both in acoustic and language modeling, neural networks today mark the state of the art for large vocabulary continuous speech recognition, providing huge improvements over former approaches that were solely based on Gaussian mixture hidden markov models and count-based language models. We give an overview of current activities in neural network based modeling for automatic speech recognition. This includes discussions of network topologies and cell types, training and optimization, choice of input features, adaptation and normalization, multitask training, as well as neural network based language modeling. Despite the clear progress obtained with neural network modeling in speech recognition, a lot is to be done, yet to obtain a consistent and self-contained neural network based modeling approach that ties in with the former state of the art. We will conclude by a discussion of open problems as well as potential future directions w.r.t. to neural network integration into automatic speech recognition systems.