Some experiments with a linguistic processor for continuous speech recognition

The linguistic processor of a system for the recognition of fluently spoken Japanese is described. Input to the processor is a phoneme lattice in which scores are given for each of 27 phonemes for each segment. Phrases composed of words from a 112 word vocabulary are recognized by an error correcting parser which tolerates both classification and segmentation errors in the phoneme lattice by means of a Substitution-Insertion-Deletion Mechanism (SID). Some phrase errors are corrected at the sentence level by means of a Cartesian Product Sorting Algorithm. The processor has been tested on a total of 80 sentences from four male speakers. The sentences comprised 496 phrases with an average of 25 phonemes per phrase. The phoneme lattices had a 70 percent accuracy on phonemes with roughly equal numbers of segmentation and classification errors. Under these conditions 77 percent phrase accuracy was observed.