Abstract

In this paper we describe a technique for non-keyword
rejection and we will evaluate in the context of an
audiotex service using the ten Spanish digits. The
baseline keyword recognition system is a
speaker-independent continuous density Hidden Markov
Model recognizer. We propose the use of an affine
transformation to the log-probability of the garbage
model, an HMM model trained to account for both nonkeyword
speech and non-stationary telephone noises. The
parameters of the transformation for the case of isolated
keywords are chosen to minimize a cost function that
weighs the keyword error rate, keyword rejection rate
and false acceptance rate according to the a priori
probabilities of keywordhon-keyword and the
requirements of the specific application. This technique
was also extended to embedded keywords (word-spotting).
Use of this rejection technique on the audiotex
application reduced the total cost function up to 20% for
isolated-word case and 12% for the word-spotting case.