chapter  10
31 Pages

Model Compensation and Matched Condition Methods for Robust Speech Recognition

WithRita Singh, Bhiksha Raj, Richard M. Stern

In this chapter, the authors describe some important model compensation and matched condition techniques used in speech recognition systems. Some key model compensation techniques are model decomposition, parallel model combination (PMC), maximum likelihood linear regression, maximum a posteriori (MAP) adaptation, and extended maximum a posteriori adaptation. Most of the compensation methods assume that the recognizer uses hidden Markov models (HMM) to model speech. Since most model compensation and matched condition techniques take into consideration the specific model used by the recognizer, the authors begin with a brief description of HMMs. In model decomposition, corrupting noises and clean speech are both modeled explicitly using HMMs. In PMC, the distributions of the clean speech are modified to approximate the distributions that optimally represent the noisy test speech. In MAP adaptation, model parameters are reestimated based on examples of the noisy data and prior knowledge of the distributions of the parameters.