Relationships between phoneme recognition accuracy and sentence recognition accuracy

Abstract

The relationships between phoneme recognition accuracy and sentence recognition accuracy for continuous speech recognition using a simulated phoneme recognizer are described. A left‐to‐right and top‐down parser based on Earley's algorithm was used as a syntactic parser. The beam search technique was embedded in this parser to reduce the search space. The syntactic constraints were represented by a context‐free grammar. The word lattice for an utterance was generated by a word spotting algorithm from an ambiguous phoneme sequence. The input to the parser was the word lattice of spotted words that consisted of a group of four terms: word name, beginning point, ending point. and matched score. Three tasks, “connected word (digits or city names),” “computer network (about 250 words),” and “inquiries on UNIX (about 500 words)” were used for the evaluation. From experiments, the conclusions are as follows. (1) It is difficult to recognize continuous speech without syntactic knowledge, (2) The search space ought to be large for managing an ambiguous input. (3) The sentence recognition rate is not so sensitive to vocabulary size.