Word-Juncture Modeling for HMM-Based Continuous Speech Recognition

It is well-known that, in general, a word's pronunciation is different when the word is uttered in isolation from when it is uttered in continuous speech. The main reason for such differences lies in the coarticulation phenomena arising at the boundary between adjacent words. Such pronunciation differences can cause errors in a recognition system if each word in the dictionary is phonetically transcribed from isolated pronunciations. Most contextual changes in pronunciation occur at the initial and final phones, i.e., the boundary phones. Thus, to deal with this problem, a good representation of the two-phone region across word boundaries, the word juncture region, must be provided to the recognizer.