Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Fast Likelihood Computation for Continuous-Mixture Densities Using a Tree-Based Nearest Neighbor Search

Frank Seide

Philips GmbH Forschungslaboratorien, Aachen, Germany

The Philips automatic train timetable information system AIS provides accurate information about train connections between more than 1100 German cities over the telephone [2]. Its speaker-independent speech recognizer is monophone-based and uses continuous-mixture densities. Most of the CPU time is spent on log-likelihood computation. For realtime operation, the number of densities had to be limited, sacrificing accuracy. To overcome this restriction, we developed a fast hierarchical within-mixture nearest neighbor search with logarithmic computational effort. The method degrades recognition accuracy by roughly 2-7% rel., but on the other hand allows for a larger number of densities to be processed. With the new method, the AIS log-likelihood computation was accelerated by a factor of nine retaining optimal accuracy.

Full Paper

Bibliographic reference.  Seide, Frank (1995): "Fast likelihood computation for continuous-mixture densities using a tree-based nearest neighbor search", In EUROSPEECH-1995, 1079-1083.