On the use of proximity indexes for isolated word recognition

An isolated word recognition system with a small dictionary may perform sequencial search, that is, compare the query utterance with each and every single one in the dictionary. However, when dealing with dictionaries made out of thousands of words, sequencial search is no longer a valid strategy since we would end up with such a slow recognition system that would surely be of no practical use. One approach to solve this problem is the use of proximity indexes so that the most similar word can be found quickly. The question that inmediately arises is: What proximity index is the ideal one for our purpose?. In this paper we compare several proximity indexes in their ability to find the most similar word to the query utterance among those included in the dictionary using less time for that purpose. According to our experiments a permutant-based index requires less distance evaluations at searching to find the nearest neighbor so it is faster than the pivot-based indexes included in our experiments such as the Fixed Query Array (FQA), the Burkhard-Keller Tree (BKT), the Fixed Query Tree (FQT) and the Fixed Height Fixed Query Tree (FHFQT).