Bayesian joint-sequence models for grapheme-to-phoneme conversion

We describe a fully Bayesian approach to grapheme-to-phonemeconversion based on the joint-sequence model (JSM). Usually, standardsmoothed n-gram language models (LM, e.g. Kneser-Ney)are used with JSMs to model graphone sequences (joint graphemephonemepairs). However, we take a Bayesian approach using ahierarchical Pitman-Yor-Process LM. This provides an elegant alternativeto using smoothing techniques to avoid over-training. Noheld-out sets and complex parameter tuning is necessary, and severalconvergence problems encountered in the discounted Expectation-Maximization (as used in the smoothed JSMs) are avoided. Everystep is modeled by weighted finite state transducers and implementedwith standard operations from the OpenFST toolkit. Weevaluate our model on a standard data set (CMUdict), where it givescomparable results to the previously reported smoothed JSMs interms of phoneme-error rate while requiring a much smaller training/testing time. Most importantly, our model can be used in aBayesian framework and for (partly) un-supervised training.