Improved pronunciation modelling by inverse word frequency and pronunciation entropy

We propose a new approach to rank the potential pronunciations for each word by their pronunciation frequency and inverse word frequency (pf-iwf) weights. The pronunciation set obtained in this way can then be pruned with different criteria. This approach not only considers the frequencies of occurrence of the pronunciations, but tries to minimize the extra confusion which may be introduced by pronunciation variations, such that the best overall performance can be achieved. A new entropy-based approach for pruning the pronunciation variations is also proposed. Experimental results showed that the proposed approach can not only improve the recognition performance, but make the performance more stable and less sensitive to various parameters, factors and options including the different pruning criteria. All the experiments were performed with the LDC Mandarin Call Home corpus, although the approaches and principles are definitely not limited to Mandarin Chinese.