Low-Resource Speech Recognition of 500-Word Vocabularies

Abstract

We describe techniques for enhancing the accuracy, efficiency and features of a low-resource, medium-vocabulary, grammar-based speech recognition system. Among the issues and techniques we explore are front-end speech / silence detection to reduce computational workload, the use of the Bayesian information criterion (BIC) to build smaller and better acoustic models, the minimization of finite state grammars, the use of hybrid maximum likelihood and discriminative models, and the automatic generation of baseforms from single new-word utterances. We report WER figures throughout, as appropriate.