Grapheme to Phoneme Conversion for Out of Vocabulary Words

32 concepts27 papers

When a text-to-speech (TTS) or speech recognition system processes language, it relies heavily on a pronunciation dictionary (a lexicon) to map written words to their spoken sounds. But language is alive and constantly evolving. New words emerge, proper names appear without warning, and technical jargon proliferates.

When a system encounters an Out-of-Vocabulary (OOV) word—such as a newly coined term or an uncommon surname—static dictionaries fail. If the system defaults to a naive, letter-by-letter reading, the output becomes robotic and unintelligible. To solve this, speech systems must possess a generative understanding of the relationship between orthography (spelling) and phonology (sound). This task is known as Grapheme-to-Phoneme (G2P) conversion.

The core challenge of G2P is that spelling-to-sound relationships are rarely one-to-one, especially in irregular languages like English. To build intuition for how researchers have solved this problem, we can look at three evolving strategies: latent alignment, structural priors, and modern representation learning.


The Latent Alignment Problem

Before a model can learn how to pronounce unseen words, it must first understand how known words are constructed. If we want to train a model on a word like pharaoh /fɛɹoʊ/, the model needs to discover that the letters ph map to the sound /f/, a maps to /ɛ/, and raoh maps to /ɹoʊ/. Because these boundaries are not explicitly marked in dictionaries, finding them is a latent alignment problem.

Early data-driven approaches relied on manual bootstrapping. For instance, Vincent Pagel et al. in Letter to sound rules for accented lexicon compression utilized a "hand-seeded" alignment process to initialize decision trees, allowing the system to compress massive dictionaries and generalize to unknown spelling patterns.

To eliminate manual labor, researchers turned to unsupervised statistical alignment. In Probabilistic deduction of symbol mappings for extension of lexicons, Bhiksha Raj et al. introduced an Expectation-Maximization (EM) framework that automatically deduces unidirectional mappings between $n$-tuples of characters and phonemes. By treating the alignment as a hidden variable, the model iteratively refines its counts until it discovers the most plausible phonetic boundaries.

Spelling:    p  h  a  r  a  o  h
             \ /   |  \______/
Phonemes:     f    ɛ     ɹoʊ

As web text and voice search expanded, models had to handle increasingly noisy OOV data. Keigo Kubo et al. addressed this in Evaluation of many-to-many alignment algorithm by automatic pronunciation annotation using web text mining by refining parameter estimation with a "city block distance" metric. This allowed the alignment algorithm to gracefully handle irrelevant character-to-sound mappings and deletions, making it possible to automatically mine accurate word-pronunciation pairs directly from the web.


Leveraging Structural and Morphological Priors

While statistical alignment works well for local character clusters, it often misses the broader linguistic context. Humans do not just look at letters sequentially; we recognize prefixes, suffixes, and root words.

In Analysis of Unknown Words through Morphological Decomposition, J. van de Plassche et al. demonstrated that breaking an unknown word down into its constituent morphemes allows a system to infer both its grammatical role and its pronunciation. If a system knows how to pronounce construct and understands the suffix -ion, it can synthesize the OOV word construction without ever having seen it in a training dictionary.

Taking this structural intuition further, Helen M. Meng et al. developed a hierarchical, bi-directional system in Phonological Parsing for Bi-directional Letter-to-Sound/Sound-to-Letter Generation. Instead of a "flat" mapping, their probabilistic parser navigates multiple layers of linguistic structure simultaneously—from morphology and syllable stress down to individual phonemes. This structured approach achieved high accuracy while requiring 20 times fewer parameters than flat models, proving that linguistic priors act as a powerful regularizer.

      [ Word: Unseen ]
             |
     [ Morphological Layer ]  --> Identifies prefixes/roots
             |
       [ Stress Layer ]       --> Determines vowel reduction
             |
     [ Phonetic Layer ]       --> Generates final phonemes

Structure also manifests at the macro level through word origins. The letters ch are pronounced differently in English words of Greek origin (character) versus French origin (chef). In A More Discerning and Adaptable Multilingual Transliteration Mechanism for Indian Languages, Anil Kumar Singh 0001 showed that first classifying a word's origin (e.g., native vs. foreign) allows a system to apply specialized phonetic mapping strategies, nearly doubling the performance of baseline OOV transliteration systems.


From Linear Classifiers to Deep Representation Learning

As the field shifted from rule-based systems to machine learning, early data-driven models like decision trees Kevin Lenzo et al. often struggled with overfitting when trained on limited dictionaries.

To make models more robust to noisy, real-world OOV distributions, Keigo Kubo et al. adapted the Adaptive Regularization of Weight Vectors (AROW) algorithm for G2P conversion in Grapheme-to-phoneme conversion based on adaptive regularization of weight vectors. By maintaining a confidence measure for each weight parameter, AROW updates its weights aggressively when encountering clear errors but conservatively when dealing with noisy data. The authors expanded on this in Structured soft margin confidence weighted learning for grapheme-to-phoneme conversion by introducing Structured Soft Margin Confidence Weighted (SMCW) learning, which updates multiple pronunciation hypotheses simultaneously to ensure the model generalizes well to unseen terms.

With the advent of deep learning, models gained the ability to capture complex, non-linear interactions between distant letters and phonemes. Kheang Seng et al. proposed a two-stage neural network in Letter-to-phoneme conversion based on two-stage neural network focusing on letter and phoneme contexts that explicitly models both the orthographic context of the letters and the historical context of the generated phonemes to resolve local ambiguities.

Stage 1: [ Letter Context ]  --> Captures surrounding spelling patterns
                                       |
                                       v
Stage 2: [ Phoneme Context ] --> Resolves ambiguity based on preceding sounds

In logographic languages like Chinese, the OOV problem manifests as polyphone disambiguation—determining the correct pronunciation of a character based on context. Traditional models treat pronunciations as discrete, independent "one-hot" labels. In Label Embedding for Chinese Grapheme-to-Phoneme Conversion, Eunbi Choi et al. reimagined this by mapping pronunciations into a continuous, dense embedding space using a triplet loss function. This allows the model to leverage shared phonetic properties across different characters, significantly improving its ability to handle rare or unseen tokens.

Today, state-of-the-art G2P performance relies on large sequence-to-sequence architectures, but deploying these models in production speech systems requires balancing accuracy with computational efficiency. In Token-Level Ensemble Distillation for Grapheme-to-Phoneme Conversion, Xu Tan et al. combined the representational power of the Transformer with token-level ensemble distillation. By using a diverse ensemble of large "teacher" models to generate soft phonetic targets over massive amounts of unlabeled text, they trained a compact "student" Transformer. This approach achieved a significant reduction in word error rates on benchmarks like CMUDict while shrinking the model size by six times, making robust OOV pronunciation prediction fast enough for real-time applications.

Key Foundational Work

Analysis of Unknown Words through Morphological Decomposition (1991) — contributed 3 concept elements to this direction.
Phonological Parsing for Bi-directional Letter-to-Sound/Sound-to-Letter Generation (1994) — 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_out_of_vocabulary_word_pronunciation_f9135e33,
  title = {out-of-vocabulary word pronunciation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/out-of-vocabulary-word-pronunciation/rdr_f9135e33e8c74ea7ab9f6a94c7ceb334},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_out_of_vocabulary_word_pronunciation_f9135e33_v1,
  title = {out-of-vocabulary word pronunciation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/out-of-vocabulary-word-pronunciation/rdr_f9135e33e8c74ea7ab9f6a94c7ceb334/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake