Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


Using a Generative Grammar to Train a Probabilistic Language Model for Speaker-Independent Speech Recognition

R. Cremonini, M. Ferretti, M. C. Galimberti, Giulio Maltese, Federico Mancini

IBM Semea Rome Scientific Center, Roma, Italy

This article describes a voice-activated flight information and reservation system. The system is based on the speaker-dependent, real-time, large vocabulary speech recognizer developed at the IBM Rome Scientific Center. To remove the speaker-dependency constraint, a speaker-independent acoustic model was built. To achieve a high recognition rate, a technique was developed to introduce grammatical constraints into the probabilistic language model of the recognizer. In particular, instead of substituting the language model for a grammar, we trained it on the "complete" corpus of the language to recognize. The corpus was generated using a generative grammar. No grammar was used during the recognition process. The high recognition rate attests the validity of the approach.

Full Paper

Bibliographic reference.  Cremonini, R. / Ferretti, M. / Galimberti, M. C. / Maltese, Giulio / Mancini, Federico (1991): "Using a generative grammar to train a probabilistic language model for speaker-independent speech recognition", In EUROSPEECH-1991, 399-402.