An expert system for speech spectrogram reading

One of the major problems that has plagued speech processing at the acoustic-phonetic level is the extreme variability of the speech signal. Experienced spectrogram readers presently seem to come closer to achieving acoustic-phonetic identification than do automatic techniques. For this reason, and conscious of the difficulty of gathering human expertise, we have chosen to use an expert system to formalise and test knowledge in spectrogram reading. Two aspects of expertise have been explored: knowledge and strategy. Our expert´s specific knowledge is formalised in the form of production rules and thus can be progressively modified. This knowledge covers acoustic, phonetic, and phonotactic information. A forward-chaining inference engine with variables is implemented for the control structure. It uses a global progressive strategy which manages confidence coefficients. We describe a task designed to observe system functioning and evaluate knowledge presently used. Results show that sufficient knowledge has been collected to correctly carrry out the task.