Noise-Robust Speech Recognition Using Deep Neural Network

Abstract

This thesis addresses the noise robustness of the recently developed Deep Neural Networks (DNNs) based speech recognition systems. Five techniques have been proposed. Firstly, a Mean Variance Normalization technique was developed to integrate noise statistics using Vector Taylor Series. Secondly, a Deep Split Temporal Context system was proposed to separately model sub-contexts of a long temporal span of speech. Thirdly, we revisited the missing feature theory and developed a DNN-based spectral masking system. Fourthly, an Ideal Hidden-activation Mask (IHM) was proposed to remove noise-prone latent detectors. Lastly, a noise code technique was developed to simulate IHM with reduced computational costs. Improved noise robustness has been obtained using the proposed techniques on benchmark tasks, Aurora-2 and Aurora-4. The spectral masking approach successfully ranks first in the literature on both tasks at the time of writing and is one of the most promising noise-robust techniques for DNNs.