Improvements in the BYBLOS Continuous Speech Recognition System

Abstract : The objective of this research was to develop accurate mathematical models of speech sounds for the purpose of large-vocabulary continuous speech recognition. The research focussed on three areas: developing better speech models to improve recognition accuracy, exploring new techniques for speaker-independent training, and developing speaker adaptation techniques that allow system use with a minimum of training. The work was performed within the BBN BYBLOS speech recognition system, which is based on the use of phonetic hidden Markov models. As a result of several model improvements, we have succeeded in decreasing the word rate by a factor of four for speaker-dependent and speaker- independent recognition. In speaker-independent recognition, we developed a new training paradigm in which we record speech from only a dozen speakers instead of the traditional approach of recording more than a hundred speakers. The same approach has been shown to be useful for effective speaker adaptation with only two minutes of speech training.