Equivalent node-based speech grammar optimization

Min Zhang, Cuntai Guan, Haizhou Li

Abstract: In a domain-specific speech application, for example in a computer-telephony dialogue system, a speech grammar is used to specify the words and patterns of words to be understood by a speech recognizer. The size of the recognizer search space depends on the speech grammar perplexity. In this paper, we propose a novel equivalent node-based speech grammar optimization algorithm. The proposed algorithm can optimize the grammar-based search graph by iteratively combining the equivalent nodes and their fan-in & fan-out transitions (arcs) if local properties show that this does not change the grammar coverage. Experiments on a 70k Resident Identity Numbers show that the proposed method can reduce 80% nodes and transitions, and the recognition is also found to be 50% faster without affecting accuracy. The algorithm is proven to be high effective and efficient.