Abstract

Two discriminative techniques are described (and evaluated) for estimating the parameters of the Gaussians in a large vocabulary speech-recognition system. The first technique is based on using a modification of the MMI objective function, and appears to provide no improvement over standard ML estimation. The second technique is based on a heuristic correction of the Gaussian parameters, and is seen to give a 2-5% improvement over ML estimation. 2 INTRODUCTION One common feature of many speech-recognition systems is that they are based on statistical methods i.e., the acoustic observations are modelled by probability-density-functions (pdf's), whose parameters are estimated statistically, from large amounts of training data. Typically, maximum-likelihood (ML) estimation is used; i.e., given the correct lexical transcription of the training data, the objective of the estimation is to maximize the likelihood of the training observations (feature vectors), conditioned on the correct lex...

Citations