Large vocabulary continuous Mandarin speech recognition using finite state machine

Abstract

The finite state transducer (FST), popularly used in the natural language processing (NLP) area to represent the grammar rules and the characteristics of a language, has been extensively used as the core in large vocabulary continuous speech recognition (LVCSR) in recent years. By means of FST, we can effectively compose the acoustic model, pronunciation lexicon, and language model to form a compact search space. In this paper, we present our approach to developing a LVCSR decoder using FST as the core. In addition, the traditional one-pass tree-copy search algorithm is also described for comparison in terms of speed, memory requirements and achieved character accuracy.