REDUCING SEARCH COMPLEXITY IN LOW PERPLEXITY TASKS

Abstract

In this paper we present a new method for improving the throughput of an asynchronous stack search based speech recognition system in the low perplexity applications. The algorithm reduces the acoustic fast match use in the cases where the word context information represented by the language model is sufficient to provide a reliable list of word candidates for the detailed match processing. The proposed technique improves the throughput of the system by reducing the number of fast match calls and by shortening the list of candidate words to be processed by the detailed match. Tested on the set of 3400 sentences, the new method reduces the CPU requirements of the search part of the speech recognition system by 47.7%, increasing the throughput of the entire speech system by 30.8% without degrading the recognition accuracy.