A comparison of lexicon-building methods for subword-based speech recognisers

A comparison of different algorithms for training of pronunciation dictionaries for use with subword-based speech recognisers is given. An extension to existing sub-optimal solutions is presented, and is shown to give results close to the maximum likelihood solution. The DARPA Resource Management (RM) database was used for evaluating the lexicon-building algorithms. When compared to the initial lexicon derived from the DARPA RM-distribution, improvements of recognition rates have been obtained for all lexicons trained with the different criteria. The maximum likelihood solution resulted in an 11.5% reduction in word error rate, compared to the 10.5% reduction offered by the proposed sub-optimal method