Experiments with a tree-search method for converting noisy phonetic representation into standard orthography

A tree-search algorithm was utilized in decoding the noisy phonetic output of an acoustic processor (AP) for purposes of automatic recognition of continuous speech. A 250-word lexicon and a finite-state grammar specify the tree to be searched. The search is performed in a best-first manner where partial paths having the highest similarity measure are extended first. The similarity measure of a word represents the degree of match of the noisy input with the best-fitting phonetic variant of that word. Phonetic variants for each word are generated automatically by a set of phonological rules. Computation of the similarity measure is based on statistics characterizing the performance of the AP. These statistics were obtained from training data by an iterative procedure. Performance results are given for both training and test data. Substantial improvement over earlier performance on the same data was realized.