Applying Many-to-Many Alignments and HMMs to Letter-to-Phoneme Conversion

Letter-to-phoneme conversion generally requires aligned training data of letters and phonemes. Typically, the alignments are limited to one-to-one alignments. We present a novel technique of training with many-to-many alignments. A letter chunking bigram prediction manages double letters and double phonemes automatically as opposed to preprocessing with fixed lists. We also apply an HMM method in conjunction with a local classification model to predict a global phoneme sequence given a word. The many-to-many alignments result in significant improvements over the traditional one-to-one approach. Our system achieves state-of-the-art performance on several languages and data sets.