Combined optimisation of baseforms and subword models for an HMM based speech recogniser

In this paper a framework for combined optimisation of baseforms and subword models for a speech recogniser is proposed. Given a set of subword Hidden Markov Models (HMMs) and a set of utterances of a specific word, the modified tree-trellis algorithm and the Baum-Welch re-estimation procedure is used iteratively to achieve a combined optimisation of baseforms and subword models. The DARPA Resource Management (RM) database was used to evaluate the combined optimisation scheme. The proposed method resulted in a monotonic increase in the likelihood score of both test- and training data. When compared to the initial lexicon derived from the DARPA RM-distribution and a set of initial HMMs, a 13% reduction in word error rate is achieved at best.