Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Utterance Verification for the Numeric Language in a Natural Spoken Dialogue

Mazin Rahim

AT\&T Labs - Research, Florham Park, NJ, USA

The defines the set of words or phrases that are relevant to the task of understanding and interpreting number strings, such as credit cards, telephone numbers, zip codes, etc. This paper addresses the issues involved in designing an utterance verification system for the numeric language in a natural spoken dialogue. The objective is to associate each utterance with a verification measure that reflects the confidence in recognizing its numeric language. A hierarchical mixture-of-experts is used to verify whether an utterance, being encoded using different features, is correctly recognized and interpreted by the understanding module. Experimental results on a field-trial study conclude that the proposed verification system outperforms the classical likelihood ratio distance and does significantly better than a standard Gaussian mixture model.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Rahim, Mazin (1999): "Utterance verification for the numeric language in a natural spoken dialogue", In EUROSPEECH'99, 57-60.