ETRW on Speech Processing in Adverse Conditions Cannes-Mandelieu, France November 10-13, 1992

A Robust Discrimination Method based on Selectively Trained Neural Networks

Yolande Anglade (1,2), Dominique Fohr (1), Jean-Claude Junqua (3)

(1) CRIN-CNRS & INRIA Lorraine, Vandoeuvre-lès-Nancy, France
(2) SOLLAC, Florange, France
(3) Speech Technology Laboratory, Division of Panasonic Technologies, Inc., Santa Barbara, CA, USA

The purpose of this work is to improve the automatic speech recognition of confusable words. The database considered is the American-English alphanumeric vocabulary. Our study proposes a new method using artificial neural networks and reports a comparison with a global method using hidden Markov models (HMM). The new method is based on the search for discriminative frames which bear the distinction between the confusable words. The tests, conducted on normal speech and Lombard speech with and without additive noise, show a general improvement of the recognition accuracy.

Full Paper

Bibliographic reference.  Anglade, Yolande / Fohr, Dominique / Junqua, Jean-Claude (1992): "A robust discrimination method based on selectively trained neural networks", In SPAC-1992, 175-178.