Phonetic search using an anchor-based algorithm

The research presented here focuses on a Phonetic Search method used in long speech utterances using a very large vocabulary of words. Our main goal was to reduce the search space dimension and thus decrease the overall computational complexity without adversely affecting the recognition performance. The suggested method was compared to the Exhaustive Search method in which all possible hypotheses are considered. In order to reduce this search space, we developed a parametric algorithm based on anchor points which significantly reduces the number of hypotheses. Two types of anchors points were considered: A-Priori Anchor Points, which are preselected; and Dynamic Anchor Points, which are determined during the search process itself. Four different mechanisms for determining anchor points were tested and compared to the Exhaustive Search method using over 2000 utterances from the IBM Voicemail Databases and a 100k word lexicon. We show that all mechanisms lead to a more targeted and efficient search and to a substantial reduction in the computational complexity while maintaining recognition performance.