Elamur Research⌕
☰
home / papers / 2007 / Applying Many-to-Many Alignments and Hidden Markov Models to Letter-to-Phoneme Conversion

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

Summary

The paper addresses letter-to-phoneme (L2P) conversion, which predicts a pronunciation from a written word and is a core step in speech synthesis. Training such systems requires letter–phoneme alignments, and earlier work usually forced these to be one-to-one. That choice handles letter pairs like "ph" → [f] or single letters like "x" → [k s] poorly. The authors build an EM-based many-to-many aligner by extending forward-backward training of a stochastic transducer. Letters and phonemes are grouped into chunks of up to two symbols. A letter-chunking bigram step then decides how words split into units, so no hand-made merge lists are needed. An instance-based classifier (IB1 in TiMBL) proposes candidate phonemes locally, and an HMM selects the best overall sequence. The system was tested with 10-fold cross-validation on CMUDict, English, Dutch and German Celex, and French Brulex. Many-to-many alignment raised word accuracy by 2.7–7.6 points over one-to-one alignment, and the HMM added a small further gain. The fully automatic system beats Black et al. (1998) on CMUDict and German Celex.

Key points

  • Best configuration (M-M+HMM) word accuracy:
    • CMUDict: 65.6
    • English Celex: 83.6
    • Dutch Celex: 91.4
    • German Celex: 89.8
    • French Brulex: 90.9
  • The one-to-one baseline scored between 60.3 and 87.0 on the same corpora.
  • Aligner settings: substrings were capped at length 2 on both sides (maxX = maxY = 2). Letters could map to nothing, but phonemes always needed a letter source.
  • The authors' own one-to-one aligner, which disallows null letters, beat the one-to-one alignments supplied by the PRONALSYL challenge on every corpus.
  • On top of one-to-one alignments, the HMM post-processor improved results consistently. By contrast, the CSInf method of Van Den Bosch and Canisius (2006), which decodes via constraint satisfaction, lowered accuracy on Brulex (86.5 vs 87.0), mostly because of trigram confusions.
  • The gains over Black et al. (1998) came without any handcrafted letter–phoneme lists, which that earlier system relied on.

Interesting ideas

  • I1. Learned chunk inventory: The aligner discovers multi-symbol units like "ng", "ti" or [k s] from data, so no one has to list them in advance. The same trick could suit other transliteration or normalization tasks between scripts with irregular correspondences, such as Ukrainian–Latin romanization.
  • I2. Asymmetric deletions: Silent letters are allowed, but every phoneme must come from some letter. Building such domain asymmetries into the edit operations keeps alignments meaningful and avoids awkward null inputs at prediction time.
  • I3. Local classifier plus sequence decoder: A per-position classifier proposes a short list of candidates, and an HMM picks the most coherent sequence. This is a cheap way to add output-sequence consistency to any token-level predictor without retraining it.
  • I4. Alignment as a separate, measurable component: Changing only the aligner, with the learner held fixed, produced bigger gains than changing the decoder. This suggests auditing and improving preprocessing alignments before tuning the model.
  • I5. Evaluation style: The paper reports whole-word accuracy with mean ± standard deviation over 10 folds across several languages. This makes small differences between methods interpretable and is worth copying for small lexical datasets.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

 Institutions

🎓University of AlbertaEdmonton · Canada16 works · 2,099 citations

Cite

Cite as Sittichai Jiampojamarn, Grzegorz Kondrak and Tarek Sherif. 2007. Applying Many-to-Many Alignments and Hidden Markov Models to Letter-to-Phoneme Conversion. Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{jiampojamarn2007applying,
  title = {{Applying Many-to-Many Alignments and Hidden Markov Models to Letter-to-Phoneme Conversion}},
  author = {Sittichai Jiampojamarn, Grzegorz Kondrak and Tarek Sherif},
  booktitle = {Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference},
  year = {2007},
  url = {https://aclanthology.org/N07-1047/}
}

References

17 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David W. Aha, Dennis Kibler, and Marc K. Albert. 1991. Instance-based learning algorithms. Machine Learning, 6(1):37–66. our card
  2. Yasemin Altun, Ioannis Tsochantaridis, and Thomas Hofmann. 2003. Hidden Markov Support Vector Machines. In Proceedings of the 20th International Conference on Machine Learning (ICML-2003). find
  3. Harald Baayen, Richard Piepenbrock, and Leon Gulikers. 1996. The CELEX2 lexical database. LDC96L14. find
  4. Alan W. Black, Kevin Lenzo, and Vincent Pagel. 1998. Issues in building general letter to sound rules. In The Third ESCA Workshop in Speech Synthesis, pages 77– 80. find
  5. Michael Collins. 2004. Discriminative training methods for Hidden Markov Models: Theory and experiments with perceptron algorithms. In Proceedings of the Conference on Empirical Methods in Natural Langauge Processing (EMNLP). our card
  6. Walter Daelemans and Antal Van Den Bosch. 1997. Language-independent data-oriented grapheme-to-phoneme conversion. In Progress in Speech Synthesis, pages 77–89. Springer, New York. find
  7. Walter Daelemans, Jakub Zavrel, Ko Van Der Sloot, and Antal Van Den Bosch. 2004. TiMBL: Tilburg Memory Based Learner, version 5.1, reference guide. In ILK Technical Report Series 04-02. find
  8. Robert I. Damper, Yannick Marchand, John DS. Marsters, and Alexander I. Bazin. 2005. Aligning text and phonemes for speech technology applications using an EM-like algorithm. International Journal of Speech Technology, 8(2):147–160, June. find
  9. Arthur Dempster, Nan Laird, and Donald Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. In Journal of the Royal Statistical Society, pages B:1–38. find
  10. Yannick Marchand and Robert I. Damper. 2000. A multistrategy approach to improving pronunciation by analogy. Computational Linguistics, 26(2):195–219, June. find
  11. Yannick Marchand and Robert I. Damper. 2005. Can syllabification improve pronunciation by analogy of English? In Natural Language Engineering, pages (1):1–25. find
  12. Eric Sven Ristad and Peter N. Yianilos. 1998. Learning string-edit distance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(5):522–532. find
  13. David Sankoff and Joseph Kruskal, 1999. Time Warps, String Edits, and Macromolecules, chapter 2, pages 55–91. CSLI Publications. find
  14. Terrence J. Sejnowski and Charles R. Rosenberg. 1987. Parallel networks that learn to pronounce English text. In Complex Systems, pages 1:145–168. find
  15. Paul Taylor. 2005. Hidden Markov Models for grapheme to phoneme conversion. In Proceedings of the 9th European Conference on Speech Communication and Technology 2005. find
  16. Kristina Toutanova and Robert C. Moore. 2001. Pronunciation modeling for improved spelling correction. In ACL ’02: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, pages 144–151, Morristown, NJ, USA. Association for Computational Linguistics. find
  17. Antal Van Den Bosch and Sander Canisius. 2006. Improved morpho-phonological sequence processing with constraint satisfaction inference. Proceedings of the Eighth Meeting of the ACL Special Interest Group in Computational Phonology, SIGPHON ’06, pages 41–49, June. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/jiampojamarn-2007-applying-many-many-alignments-hidden/