Regularized feature-based maximum likelihood linear regression for speech recognition

In many automatic speech recognition (ASR) applications, maximum likelihood linear regression (MLLR), and feature-based maximum likelihood linear regression (FMLLR) are used for speaker adaptation. This paper investigates a possible generalization of FMLLR which addresses the degradation in the performance of ASR systems due to small —possibly time-varying— perturbations of the training and the testing data. We formulate the problem as a regularized maximum likelihood linear regression problem. Based on this formulation, we describe a computationally efficient algorithm for estimating the linear regression parameters which maximize the sum of the log likelihood and the negative of a measure of the sensitivity of the estimated likelihood to these perturbations. This approach does not make any assumptions about the noise model during training and testing. We present several large vocabulary speech recognition experiments that show significant recognition accuracy improvement compared to using the speaker-adapted baseline models.