The model parameters for corrupted speech can be accurately and efficiently estimated by approximating the non-linear function by a Lagrange polynomial as described in this paper. The performance of speech recognizer with Lagrange polynomial approximation based model composition was significantly improved compared to other methods.
Future work includes estimation of covariance matrix, and evaluation of the approach on different tasks and with different models, such as with context-dependent phone models and Gaussian mixtures models. Furthermore, possibilities of using other polynomial expansions will be investigated.