Hierarchical DP for word recognition

A hierarchical multi-pass process using a two level dynamic programming structure for word recognition is described. Segmentation is used to define regions within the global matrix which will be used as "local" matrices for dp in a hierarchical fashion. A global warping path and distance is refined iteratively as results from local dp are incorporated. This technique is designed for more controlled interaction between regions, effective use of hierarchical processing, and speech specific pattern matching at both levels. A related experiment to test the effectiveness of regional, phonetically oriented templates and matching parameters for word pair discrimination is summarized. Word pair specific tests discriminated without error across all speakers for ten of the most confusable pairs in a standard data base (16 speakers, 5120 trials total). The tests were found using a catalogue of region types and parameters selected for potential utility in this task.