One-Pass Search Algorithm for Continuous Speech Recognition Using Generalized LR Parsing : A CFG-Driven, Frame-Synchronous HMM-Based Approach

概要

論文の詳細を見る
In this paper, we present a novel continuous speech recognition algorithm that integrates three major technologies : (1)hidden Markov models for speech, (2)a generalized LR parser for handling context-free grammar (CFG) constraints, and (3)the one-pass search algorithm for efficient searching. We also introduce three techniques that used in the development of the algorithm : (1)LR path-merging, (2)the use of a shared tree-structured stack, and(3)LR-parser-based dynamic network generation. By means of the proposed algorithm, an optimal hypothesis can be found efficiently for a given speech signal according to a specified CFG in a frame-synchronous process. We implemented an experimental Japanese speech recognition system based on the proposed algorithm, using discrete-type context-independent HMMs without duration control. The system attained a recognition accuracy of 84.1%-88.1%, depending on the beam width. We also experimentally compared our algorithm with the following two methods : (1)the one-pass search algorithm using the finite-state approximation for a CFG, and(2)the HMM-LR algorithm. The experiments showed that the proposed algorithm attained higher accuracy when the beam width was small.
一般社団法人情報処理学会の論文
1995-05-15