Applying Many-to-Many Alignments and Hidden Markov Models to Letter-to-Phoneme Conversion

Proceedings of NAACL HLT 2007, pages 372–379, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Applying Many-to-Many Alignments and Hidden Markov Models to Letter-to-Phoneme Conversion Sittichai Jiampojamarn, Grzegorz Kondrak and Tarek Sherif Department of Computing Science, University of Alberta, Edmonton, AB, T6G 2E8, Canada {sj,kondrak,tarek}@cs.ualberta.ca Abstract Letter-to-phoneme conversion generally requires aligned training data of letters and phonemes. Typically, the align- ments are limited to one-to-one align- ments. We present a novel technique of training with many-to-many alignments. A letter chunking bigram prediction man- ages double letters and double phonemes automatically as opposed to preprocess- ing 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 tradi- tional one-to-one approach. Our system achieves state-of-the-art performance on several languages and data sets. 1 Introduction Letter-to-phoneme (L2P) conversion requires a sys- tem to produce phonemes that correspond to a given written word. Phonemes are abstract representa- tions of how words should be pronounced in natural speech, while letters or graphemes are representa- tions of words in written language. For example, the phonemes for the word phoenix are [ f i n I k s ]. The L2P task is a crucial part of speech synthesis systems, as converting input text (graphemes) into phonemes is the first step in representing sounds. L2P conversion can also help improve performance in spelling correction (Toutanova and Moore, 2001). Unfortunately, proper nouns and unseen words pre- vent a table look-up approach. It is infeasible to con- struct a lexical database that includes every word in the written language. Likewise, orthographic com- plexity of many languages prevents us from using hand-designed conversion rules. There are always exceptional rules that need to be added to cover a large vocabulary set. Thus, an automatic L2P sys- tem is desirable. Many data-driven techniques have been proposed for letter-to-phoneme conversion systems, including pronunciation by analogy (Marchand and Damper, 2000), constraint satisfaction (Van Den Bosch and Canisius, 2006), Hidden Markov Model (Taylor, 2005), decision trees (Black et al., 1998), and neural networks (Sejnowski and Rosenberg, 1987). The training data usually consists of written words and their corresponding phonemes, which are not aligned; there is no explicit information indicating individual letter and phoneme relationships. These relationships must be postulated before a prediction model can be trained. Previous work has generally assumed one-to-one alignment for simplicity (Daelemans and Bosch, 1997; Black et al., 1998; Damper et al., 2005). An expectation maximization (EM) based algo- rithm (Dempster et al., 1977) is applied to train the aligners. However, there are several problems with this approach. Letter strings and phoneme strings are not typically the same length, so null phonemes and null letters must be introduced to make one- to-one-alignments possible, Furthermore, two letters frequently combine to produce a single phoneme 372 (double letters), and a single letter can sometimes produce two phonemes (double phonemes). To help address these problems, we propose an automatic …

Content truncated to 500 words