Phone recognition for spoken web search

Abstract

Aiming at both speaker independence and robustness with respect to recognition errors in the spoken queries, we have implemented a two-pass system for spoken web search. In the first pass, unconstrained phone recognition of both the query terms and the content audio is employed to represent these recordings as phone strings. A dynamic-programming approach then finds regions in the content phone strings that correspond closely to one or more query strings. In the sec- ond pass, each of these regions is again processed with a phone recognizer, but now a lattice is extracted; this lattice is com- pared against similar lattices extracted for each of the queries. We find our approach to be somewhat successful in identify- ing the query terms in both the development and evaluation sets, but not to generalize well between these sets.