Reducing speech recognition time and memory use by means of compound (de-)composition

This paper tackles the problem of Out Of Vocabulary words in Automatic Speech Transcription applications for a compound language (Dutch). A seemingly attractive way to reduce the amount of OOV words in compound languages is to extend the AST system with a compound (de-)composition module. However, thus far, successful implementations of this approach are rather scarce. We developed a novel data driven compound (de-)composition module and tested it in two different AST experiments. For equal lexicon sizes, we see that our compound processor lowers the OOV rate. Moreover we are able to transform that gain in OOV rate into a reduction of the Word Error Rate of the transcription system. Using our approach we built a system with an 84K lexicon that performs as accurately as a baseline system with a 168K lexicon, but our system is 5-6% faster and requires about 50% less storage for the lexical component, even though this component is encoded in an optimal way (prefix-suffix tree compression)