Robust Estimation and Adaptation of Subspace Gaussian Mixture Models for Automatic Speech Recognition

In conventional hidden Markov model (HMM) based speech recognisers, the emitting HMM states are modelled by Gaussian Mixture Models (GMMs), with parameters been estimated directly from the training data. However, in Subspace Gaussian mixture model(GMM) based acoustic modelling, the parameters of each state model are derived from the globally shared model subspaces which are normally low dimensional. This leads to significant reduction in terms of total parameters to be estimated while allowing larger number of Gaussians to be used to increase the model capacity. Considerable performance gains are observed in several speech recognition tasks compared to conventional acoustic modelling. Though relatively compact, SGMM acoustic models still suffer from model overfit problem given small amount of training data. In this report, we will discuss the model estimation with regularizations to improve robustness, and model estimation from out-domain data to increase the model accuracy. We will also show that these techniques can be successfully applied to develop a target language system with low resource by cross-lingual setting. In addition, to address the mismatch between model training and testing, we will also present the adaptation techniques of SGMM acoustic model. In particular, we will show how the model estimated from outdomain data can be adapted by maximum a posteriori (MAP) criteria to fit the target system better, and also the model trained in clean condition can be adapted with noise compensation techniques to work well in noisy environment.