Vocabulary independent acoustic-phonetic modeling for continuous speech recognition

This paper investigates the problem of defining the acoustic-phonetic unit set for flexible vocabulary continuous speech recognition systems. As an alternative to the classical modeling approach with biphones and triphones, a set of stationary/transitory state units is defined that is limited enough in number as to represent a closed set trainable once and for all. A major benefit of these units is that inter-word transitions can easily be taken into account. We show that a system employing these new units favorably compares with respect to a baseline recognizer with Continuous Density Hidden Markov Models of context-dependent biphones and triphones, selected through a minimal occurrence criterion within the training database.