Large vocabulary word detection by searching in a tree-structural word dictionary

A large vocabulary word detection method is presented for use with a Japanese continuous speech recognizer. This word detector receives a syllable lattice resulting from syllable recognition of a continuously spoken Japanese word sequence, and produces word candidates and their locations in speech as output. Top-down word detection using a tree-structural word dictionary is carried out to allow the toleration of syllable recognition errors in the syllable lattice and to reduce computational cost. All word candidates at arbitrary locations in the lattice are detected by traversing the dictionary only once. In order to reduce the space explored in the dictionary, nodes whose syllable symbols are contained in the lattice are marked before traversing the dictionary. As the result of evaluation experiments for a 4000 word vocabulary, the correct word candidates were detected in the top 2.5% and 20% of the vocabulary with error rates less than 9% and 4%, respectively.