In this paper, we deal with the problem of non-keyword modelling and rejection in a Hidden Markov Model (HMM) based Spanish keyword spotting. When talking about the performance of a keyword spotting system in terms of false alarm rejection, the non-keyword modelling and the rejection techniques are two relevant topics. With regard to the non-keyword modelling, our approach is to define a set of task independent filler models which can be used in any application. In this paper we investigate the performance of a set of filler definition in the problem of detecting digits embedded in utterances. Particularly, we are working with three filler definitions: phonetic fillers, syllabic fillers and word-based fillers. For false alarm rejection, we handle the problem as a post processor of the HMM word spotting recogniser. We design a specific classifier based on a Neural Network and linear discriminant functions to classify a keyword hypothesis in keywonl/non-keyword.

Keywords: Keyword spotting, hidden Markov models, filler models, false alarm rejection, linear discriminant functions, Neural Network.