Smooth interpolation of Gaussian mixture models

The article describes an approach for embodiment of the ambient noise awareness into the statistical model of individual speech units to preserve speech recognizer´s robustness under varying environmental conditions. Unlike previous approaches, continuous model for a given range of noise parameters is suggested to allow precise modeling of any predictable ambient noise conditions. The method is based on linked expectation maximization training of a series of Gaussian mixture models in a dense SNR stepping fashion followed by averaging and decimation to reduce storage needs. Final models are interpolated using piecewise cubic Hermite polynomial preserving the shape of the initial model set.