斉藤博昭
音韻ラティスを用いての文節認識
Abstract:This paper describes an experiment in continuous speech recognition.
The task is producing phrase candidates from a phoneme lattice. Because lattice
parsing requires a much more extensive search than conventional string input
parsing, an LR parsing algorithm is adopted as an efficient parsing method. The
system is evaluated against continuous speech data using a fairly extensive
context free grammar. An 82 % recognition rate is achieved for the top five
candidates. Average parsing time is 13 seconds.