Subspace Gaussian Mixture Models for Large Vocabulary Speech Recognition

Subspace Gaussian mixture model(GMM) is an alternative approach to approximate the probabilistic density function (p.d.f) of a set of independent identical distributed (i.i.d) data with prior density estimates. In this approach, the prior density of GMM parameters is estimated from a development dataset, and when predict the new enrolled data, the prior knowledge can be utilised by criteria like Maximum a Posterior. Unlike the conventional prior estimate method for GMM, the correlations between parameters of different Gaussian components are considered in this approach. In order to handle the large size of parameter set and meanwhile to ensure the priors be informative, the prior density estimation is constraint to a low dimensional subspace of the whole model space which can capture the main model variations. The subspace GMM has already been successfully applied in the task of speaker recognition, and achieved promising performance, but there is no much work of applying this approach to speech recognition. In this paper, we will present a new framework of HMM based speech recognition system based subspace GMM, in which, the parameters of state-dependent GMM are not estimated separately but been generated from the globally shared low dimensional model subspace. The approach can considerably reduce the model size and in addition, make the speech recognition system more scalable and adaptable. In this paper, we will first review the principles of subspace GMM approach based on its applications in speaker recognition and then discuss how to extend it to the task of speech recognition.