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

Improving the Robustness of Automatic Speech Recognizers Using State Duration Information

N. Nicol (1), Stephan Euler (2), M. Falkhausen (1), Herbert Reininger (1), Dietrich Wolf (1), J. Zinke (2)

(1) Institut fiir Angewandte Physik, Johann Wolfgang Goethe-Universität, Frankfurt a.M., Germany
(2) Telenorma, Bosch Telecom, Frankfurt a.M., Germany

In this paper we discuss the influence of state duration information on the robustness of an isolated word recognition system operating in noisy environment. Two methods for modeling state durations in Hidden Markov Models are compared. First, a method for modeling the distribution of state durations with Poisson statistics and second, a method of internal state duration modeling. Recognition results obtained with a vocabulary of 23 German words disturbed with two different kinds of noise are presented.

Full Paper

Bibliographic reference.  Nicol, N. / Euler, Stephan / Falkhausen, M. / Reininger, Herbert / Wolf, Dietrich / Zinke, J. (1992): "Improving the robustness of automatic speech recognizers using state duration information", In SPAC-1992, 183-186.