An efficient phoneme distance measure using a lexical tree

The high complexity associated with phonetic search processes when using very large lexica, long utterances and huge speech databases has become a vital issue over the last decade. In this paper, we present a novel phonetic search method based on a lexical tree that addresses this problem. The new suggested method was compared to a basic phonetic search using over 2000 utterances from the IBM Voicemail Databases and a 100k word lexicon. Preliminary results show that using a phonetic search based on the suggested tree method, does not degrade the overall recognition performance and reduces computational complexity by 35%.