Integrating probabilistic LR parsing into speech understanding systems

The application of probabilistic LR parsing to the problem of continuous speech understanding is described. A probabilistic model for LR parsing is introduced, the integration of the parser into a speech understanding system is discussed, and recognition results are presented. The goal of speech understanding is to process continuous speech input from users and provide appropriate responses. This requires a speech recognition system with a high sentence recognition rate, and a low sentence error rate (since rejection is allowed, these are not identical). The authors demonstrate that probabilistic LR parsing can help meet these requirements by applying linguistic and statistical constraints to speech recognition, and by providing effective rejection criteria. In experiments using the MIT VOYAGER spontaneous speech corpus, the use of a probabilistic LR parser improved the percentage of utterances for which correct semantics were produced from 23% (using a perplexity 72 word-pair grammar) to 58%. System performance, as measured by the metric 100× (N_corret-N_error)/N_total , was 44.5 at a 32.7% rejection rate