Keywords: Speech recognition, speech understanding, N-best paradigm, syntactic natural language processing, acoustic modeling, hidden Markov models, data oriented parsing. hidden Markov models, data oriented parsing. hidden Markov models, data oriented parsing. hidden Markov models, data oriented parsing. hidden Markov models, data oriented parsing.hidden Markov models, data oriented parsing. Contact Person: Martin Rajman Phone: (+41 21) 693-5277 E-mail: Martin.Rajman@epfl.ch Partners: IDIAP
The main research goal of this project is to develop and assess new strategies for integrating state-of-the-art acoustic models and advanced linguistic models into speech understanding systems, in view of improving dialog-based interactive voice response (IVR) systems.
Within the general framework of the N-best paradigm, the research theme of the INSPECT project is to investigate new ways of generating the N-best hypotheses and to produce new sets of hypotheses that include more ``semantic'' variability, becoming therefore more appropriate for linguistic post-processing. This is a very challenging and pretty much untouched research area in which we believe we have identified promising directions to investigate. In particular, we will concentrate on the generation of better N-best hypotheses by
- using longer term acoustic knowledge and confidence levels and by
- applying methods used in robust parsing directly to the acoustic level, e.g., on phonetic strings.
The applicative framework that is used for the evaluation of the performance is the one currently developed in the ISIS Project.
This project is funded by grant FNRS #21-54100.98.