A robust connected-words recognizer

Speaker-independent recognition of connected words is considered. Robustness against variations of both background noise and frequency responses of microphones and transmission lines is improved by using a high-pass filter to reduce stationary or slowly varying parts of the spectral components in the feature vectors. A method is proposed for improving the power of references generated from isolated words by combining them with duration models derived from a small set of connected words. Results of experiments to improve recognition accuracy using speaker-dependent adaptation of the hidden Markov models´ transition probabilities and of feature vectors are presented. Results are given for a German corpus and the TI/NIST connected digits corpus