Automatic baseform generation from acoustic data

Abstract

We describe two algorithms for generating pronunciation networks from acoustic data. One is based on raw phonetic recognition and the other uses the spelling of the words and the identification of their language of origin as guides. In both cases, a pruning and voting procedure distills the noisy phonetic sequences into pronunciation networks. Recognition experiments on two large, grammar-based, test sets show a reduction of sentence error rates between 2% and 14%, and of word error rate between 3% to 23% when the learned baseforms are added to our baseline lexicons.