Advanced acoustic techniques in automatic speech understanding

The BBN Speech Understanding System (HWIM) has two ways by which it may access the acoustic data of an unknown utterance. The first is an Acoustic-Phonetic Recognition component which derives a bottom-up phonetic transcription of the utterance. New techniques in probabilistic modeling of multi-dimensional density distributions have allowed many acoustic features to be fully utilized. HWIM (for Hear What I Mean) also contains a Parametric Word Verification component which does a top-down parametric word match to determine if the acoustic evidence is consistent with the presence of a hypothesized word. The component uses a synthesis-by-rule program to generate synthetic templates and a dynamic programming algorithm to do time normalization.