An improved union model for continuous speech recognition with partial duration corruption

The probabilistic union model is improved for continuous speech recognition involving partial duration corruption, assuming no knowledge about the corrupting noise. The new developments include: an n-best rescoring strategy for union based continuous speech recognition; a dynamic segmentation algorithm for reducing the number of corrupted segments in the union model; a combination of the union model with conventional noise-reduction techniques to accommodate the mixtures of stationary noise (e.g. car) and random, abrupt noise (e.g. a car horn). The proposed system has been tested for connected-digit recognition, subjected to various types of noise with unknown, time-varying characteristics. The results have shown significant robustness for the new model.