Spanish Keyword Spotting System Based on Filler Models, Pseudo N-gram Language Model and a Confidence Measure

In order to organize efficiently lots of hours of audio contents such as meetings, radio news, search for spoken keywords is essential. An approach uses filler models to account for non-keyword intervals. Another approach uses a large vocabulary continuous speech recognition system (LVCSR) which retrieves a word string and then search for the keywords in this string. This approach yields high performance but it requires a lot of training data and costly computation. In this paper we present several filler models and a confidence measure explored in a Spanish keyword spotting system. We will also investigate different weights in the grammar used for the language modelling in the keyword spotting system in order to achieve the best results. The keyword technique used is based on Hidden Markov Model (HMM). Test results are reported on a set of data from the geographic corpus of Albayzin speech data base containing 80 keywords taken from the words which most times occurs in the corpus sentences.