A Procedure to Generate Training Sequences for a Connected Word Recognizer Using the Segmental k-means Training Algorithm

It has been shown [1] that a connected digit recognizer, based on either word templates or hidden Markov models (HMM), could effectively be trained using a segmental k-means training procedure. For such a procedure, a set of randomly generated digit strings of variable length is used to train the recognizer. However, problems are encountered when this continuous speech training procedure is extended to large vocabulary, connected word speech recognition.