Letter-Phoneme Alignment: An Exploration

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 780–788, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Letter-Phoneme Alignment: An Exploration Sittichai Jiampojamarn and Grzegorz Kondrak Department of Computing Science University of Alberta Edmonton, AB, T6G 2E8, Canada {sj,kondrak}@cs.ualberta.ca Abstract Letter-phoneme alignment is usually gen- erated by a straightforward application of the EM algorithm. We explore several al- ternative alignment methods that employ phonetics, integer programming, and sets of constraints, and propose a novel ap- proach of refining the EM alignment by aggregation of best alignments. We per- form both intrinsic and extrinsic evalua- tion of the assortment of methods. We show that our proposed EM-Aggregation algorithm leads to the improvement of the state of the art in letter-to-phoneme con- version on several different data sets. 1 Introduction Letter-to-phoneme (L2P) conversion (also called grapheme-to-phoneme conversion) is the task of predicting the pronunciation of a word given its orthographic form by converting a sequence of letters into a sequence of phonemes. The L2P task plays a crucial role in speech synthesis sys- tems (Schroeter et al., 2002), and is an important part of other applications, including spelling cor- rection (Toutanova and Moore, 2001) and speech- to-speech machine translation (Engelbrecht and Schultz, 2005). Many data-driven techniques have been proposed for letter-to-phoneme conversion systems, including neural networks (Sejnowski and Rosenberg, 1987), decision trees (Black et al., 1998), pronunciation by analogy (Marchand and Damper, 2000), Hidden Markov Models (Taylor, 2005), and constraint satisfaction (Bosch and Can- isius, 2006). Letter-phoneme alignment is an important step in the L2P task. The training data usually consists of pairs of letter and phoneme sequences, which are not aligned. Since there is no explicit infor- mation indicating the relationships between indi- vidual letter and phonemes, these must be inferred by a letter-phoneme alignment algorithm before a prediction model can be trained. The quality of the alignment affects the accuracy of L2P con- version. Letter-phoneme alignment is closely re- lated to transliteration alignment (Pervouchine et al., 2009), which involves graphemes representing different writing scripts. Letter-phoneme align- ment may also be considered as a task in itself; for example, in the alignment of speech transcription with text in spoken corpora. Most previous L2P approaches induce the align- ment between letters and phonemes with the ex- pectation maximization (EM) algorithm. In this paper, we propose a number of alternative align- ment methods, and compare them to the EM- based algorithms using both intrinsic and extrin- sic evaluations. The intrinsic evaluation is con- ducted by comparing the generated alignments to a manually-constructed gold standard. The extrin- sic evaluation uses two different generation tech- niques to perform letter-to-phoneme conversion on several different data sets. We discuss the ad- vantages and disadvantages of various methods, and show that better alignments tend to improve the accuracy of the L2P systems regardless of the actual technique. In particular, one of our pro- posed methods advances the state of the art in L2P conversion. We also examine the relationship be- tween …

Content truncated to 500 words