Linguistically Informed Grapheme to Phoneme Conversion

32 concepts27 papers1 proposal

When we read the word cat, our brains instantly map the letters c-a-t to the spoken sounds /kæt/. For a computer, this Grapheme-to-Phoneme (G2P) conversion is a complex mapping challenge. If every language possessed a perfectly transparent orthography—where one letter always corresponds to exactly one sound—G2P would be a trivial lookup table.

However, human language is rarely so simple. The pronunciation of a character shifts depending on its neighbors, its position within a syllable, lexical stress, and morphological boundaries. To resolve these ambiguities, speech synthesis and recognition systems rely on architectures that capture the underlying rules of human phonology.

The Architecture of Phonetic Rules

To explicitly instruct a machine how to pronounce a word, traditional systems construct sequential pipelines of context-dependent rules. As shown by Anderson et al. in their work on Old Irish, this process typically breaks down into three distinct stages:

  1. Preprocessing: Normalizing and preparing the raw text.
  2. Non-contextual mapping: Establishing baseline character-to-sound assignments.
  3. Postprocessing: Applying contextual adjustments to account for surrounding letters.

To handle complex phonetic phenomena like consonant cluster reductions or assimilation, systems often apply regular expressions sequentially. For instance, Senyk et al. developed a rule-based phonemizer for Ukrainian that uses sequential regular expressions to resolve local phonetic shifts.

Similarly, in languages like Greek, specific rules must be crafted to handle context-dependent variations such as diphthongs and iotacism (where different graphemes converge on the same vowel sound). This challenge was addressed by both Lyras et al. and Chalamandaris et al., demonstrating that well-mapped phonetic rule systems can achieve high accuracy on out-of-vocabulary words without requiring vast training data.

Why Local Context Fails: Syllables, Stress, and Morphology

A simple sliding window of a few characters is often insufficient for accurate G2P conversion because pronunciation is governed by higher-level linguistic structures.

Syllables and Stress

In many languages, a vowel's pronunciation depends entirely on whether it is stressed. In Italian, as noted by Savino et al., assigning lexical stress—usually to the penultimate syllable—is a prerequisite for correct phonetic mapping. Daelemans demonstrated with the GRAFON system for Dutch that a syllable-centric approach yields exceptionally high token accuracy. Even when statistical models like Hidden Markov Models (HMMs) are used, incorporating explicit syllable-level stress models drastically improves performance over baseline character models, as shown by Taylor.

Morphology

Words are composed of meaningful units called morphemes. When two morphemes meet at a boundary, standard phonetic rules often break down. For example, the "g" in finger sounds different from the "g" in hanger due to morphological structure. Taylor et al. demonstrated that explicitly marking morphological boundaries in text inputs significantly reduces word and phoneme error rates in sequence-to-sequence models, particularly when encountering unseen root morphemes.

Structural Ambiguity

In some languages, certain sounds are not written at all. In Sinhala, the system must resolve schwa epenthesis—inserting an unwritten vowel sound based on structural constraints. Wasala et al. showed that linguistically-informed rules can resolve these inherent ambiguities with near-perfect precision, offering a computationally efficient alternative to complex statistical models.

[Raw Graphemes] ──> [Morphological Segmentation] ──> [Syllabification & Stress] ──> [Contextual Rules] ──> [Phonemes]

From Handcrafted Rules to Hybrid Learning

While expert-defined rules are highly precise, they struggle with exceptions, loanwords, and polyphones (words with identical spellings but different pronunciations). For example, Beniamine et al. highlighted the extensive manual refinement and verification needed to handle loanwords in Estonian. To alleviate this cognitive load, researchers turned to machine learning to discover, optimize, or augment these rules.

Transformation-Based Learning (TBL)

Instead of writing rules by hand, TBL starts with a baseline mapping and automatically learns an ordered sequence of correction rules based on errors in a training corpus. Zheng et al. demonstrated that TBL dramatically outperforms standard decision trees for Mandarin polyphones, successfully correcting complex characters like '为'.

Statistical Machine Translation (SMT) and Alignment

G2P can also be framed as translating a sentence of letters into a sentence of sounds. Rama et al. framed G2P as a phrase-based SMT problem, showing significant accuracy gains on German and Dutch by treating sequences of letters as phrases. Crucial to this translation is flexible alignment; Jiampojamarn et al. introduced online discriminative training using many-to-many alignments to capture multi-letter phonetic mappings effectively.

Hybrid Systems

The most robust systems often combine data-driven learning with structural linguistic constraints:

  • Thai: Charoenpornsawat et al. combined an example-based approach with a rule-based syllable segmenter to outperform pure rule systems. For transcribing English loanwords into Thai, Thangthai et al. utilized a hybrid approach combining dictionary lookups, learned rules, and syllable constraints.
  • Korean: Lee et al. designed a hybrid system where statistical components handled complex phonological rules like /n/-insertion and tensification, which are notoriously difficult to hardcode cleanly.

Modern Multilingual and Neural Integration

Today, the field has moved toward deep neural networks (DNNs) and cross-lingual induction, yet linguistic constraints remain vital to prevent models from making unlinguistic errors.

When building a G2P system for a new language with limited data, we can leverage existing resources. Bleyan et al. demonstrated that finite-state transducers can induce G2P rules across languages sharing a writing system (such as Latin scripts), drastically reducing manual effort.

For high-resource scenarios, modern DNN-based G2P converters, such as the one proposed by Jůzová et al., process entire phrases rather than isolated words to capture long-range phonetic interactions. To ensure these powerful neural models remain consistent, researchers inject structural knowledge. For example, Lo et al. won the low-resource SIGMORPHON 2021 challenge by explicitly penalizing vowel errors and incorporating syllable structure into their model's loss function.

Finally, Ohnaka et al. demonstrated that combining implicit neural representations (like BERT embeddings) with explicit dictionary pruning ensures strict grapheme-phoneme consistency without sacrificing prosodic naturalness. By anchoring flexible neural architectures with rigid linguistic boundaries, modern G2P systems achieve both high accuracy and natural pronunciation.

Key Foundational Work

Unlimited Vocabulary Grapheme to Phoneme Conversion for Korean TTS (1998) — contributed 2 concept elements to this direction.
GRAFON: A Grapheme-to-Phoneme Conversion System for Dutch (1988) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_rule_based_grapheme_to_phoneme_conversio_cc84e38b,
  title = {rule-based grapheme-to-phoneme conversion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/rule-based-grapheme-to-phoneme-conversion/rdr_cc84e38b9a25497f8cef1ef4ddd8175c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_rule_based_grapheme_to_phoneme_conversio_cc84e38b_v1,
  title = {rule-based grapheme-to-phoneme conversion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/rule-based-grapheme-to-phoneme-conversion/rdr_cc84e38b9a25497f8cef1ef4ddd8175c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake