Speech Recognizer with Dynamic Alternative Path Search and Its Performance Evaluation

摘要

For a middle-size (around 1,000 words) vocabulary speech recognition, a Finite State Automaton (FSA) language model is widely used. However, defining a FSA model with sufficient coverage and consistency requires much human effort. We already proposed a method to automatically construct a FSA language model from learning corpus by use of FSA DP matching algorithm. Experiment results show that this model attains quite high recognition correct rate for closed data, but only low rate for open data. This is mainly because a necessary path does not appear in a generated FSA. To cope with this problem, we propose a new search algorithm that allows to jump dynamically to an alternative path when speech recognition of some words seems to fail. We report experiment results and discuss the effectiveness of the algorithm.