Using natural-language knowledge sources in speech recognition

The main purpose of this paper is to investigate a hybrid language model incorporating both a unification grammar and \(n\)-gram statistics, and to show that such a model may really increase the speech recognition accuracy. Probabilistic context-free grammars and probabilistic unification grammars can also provide possibly, valuable solutions. These approaches are coming from two basically different situations: (1) When training data for statistical models is not available, methods for using complex natural-language grammars by compiling them into context-free grammars that can be efficiently incorporated into the search architecture of a commercially available real-time speech recognizer. (2) When adequate training data for statistical models is available, the author describes an experiment in which a hybrid language model incorporating both \(n\)-gram statistics and constraints from a qualitative natural-language grammar substantially improved the performance of a conventional hidden Markov model-based recognizer. Finally, statistical context-free grammars are proved to possibly incorporate all the constraints of a complex unification-based grammar. For the entire collection see [Zbl 0927.00044].

Recommendations

Cited in

(8)

This page was built for publication: Using natural-language knowledge sources in speech recognition

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4942754)