Grapheme to Phoneme Conversion for Proper Names

23 concepts16 papers1 proposal

Standard text-to-speech (TTS) and automatic speech recognition (ASR) systems rely on Grapheme-to-Phoneme (G2P) conversion to map written letters into spoken sounds. For standard vocabulary, this process is highly effective. In languages with regular orthographies like Greek, A. Chalamandaris et al. show that rule-based systems can achieve over 98% accuracy on unseen words. Even in more complex languages, systems can leverage morphological analysis to break down compound words into predictable units, as demonstrated by Amanda Pounder et al., or use statistical frameworks like Hidden Markov Models, as explored by Paul Taylor, to capture context-sensitive pronunciation rules.

However, proper names—such as people, places, and brands—routinely break these rules.

Why Names Break Standard G2P

A proper name is not just a collection of letters; it is a linguistic artifact that carries its own historical and geographical context. When a foreign name is introduced into a language, it often retains its original pronunciation while keeping its native spelling. Because standard G2P models are trained on general vocabulary, they fail to anticipate these etymological exceptions.

As Philippe Boula de Mareuil et al. found, even state-of-the-art G2P systems suffer from error rates of 12% to 20% when processing proper names. This performance drop creates a significant bottleneck for speech technologies. Willem D. Basson et al. observed that the accuracy gap between simple grapheme-based ASR and phoneme-based ASR is heavily concentrated within specific categories like proper names and loanwords. When a system encounters an out-of-vocabulary name, traditional data-driven techniques struggle to generalize, though objective evaluations by Tasanawan Soonklang et al. show that "pronunciation by analogy" can occasionally mitigate these errors better than standard rule lookups.

To solve this, researchers have developed specialized strategies that move beyond generic letter-to-sound rules.

Strategy 1: Conditioning on Origin and Cross-Lingual Hints

If names carry their linguistic history with them, a G2P system should explicitly model that history. One intuitive approach is to identify a name's country of origin before attempting to pronounce it. Sonjia Waxmonsky et al. implemented this by combining language-specific pronunciation models with a general model, using word-origin classifiers to weight the phonetic output.

But what if the origin of a name is ambiguous? Aditya Bhargava et al. introduced a method that uses cross-lingual transliterations as phonetic "hints." For example, if an English name has an established transliteration in a phonetically transparent script (like Russian or Hindi), that secondary spelling provides explicit clues about vowel quality and syllable stress. By re-ranking G2P hypotheses using these cross-lingual transliterations, the system resolves orthographic ambiguities that are impossible to untangle from the English spelling alone.

[Written Name] ───> [Cross-Lingual Transliteration (e.g., Cyrillic)] ───> Phonetic Hint
       │                                                                       │
       ▼                                                                       ▼
[Standard G2P Engine] ───────────────────────────────────────────────> [Discriminative Re-ranking] ───> Correct Phonemes

Strategy 2: Phoneme-to-Phoneme Post-Processing

Rather than rebuilding a core G2P engine to accommodate every global naming convention, an alternative approach splits the task into a two-stage pipeline: a naive guess followed by a targeted correction.

Henk van den Heuvel et al. designed a system where a general-purpose G2P converter first generates a baseline phonetic transcription. This baseline is then passed to a specialized Phoneme-to-Phoneme (P2P) converter trained exclusively on proper names. The intuition is that a general G2P engine will make predictable, systematic errors when encountering foreign names. The P2P model learns to recognize these systematic errors and translate them into the correct target phonemes.

This architecture also helps capture speaker variation. Henk van den Heuvel et al. later extended this mechanism to generate multiple pronunciation variants. This allows ASR systems to successfully recognize foreign names whether they are pronounced with their original native phonetics or adapted to local accents.

Strategy 3: Grounding in Audio and Crowdsourcing

When text-based rules and statistical priors fail, the most reliable source of truth is empirical evidence: how people actually pronounce the name in conversation.

One way to capture this is by analyzing speech recognition failures. Sravana Reddy et al. developed an Expectation-Maximization (EM) framework that expands pronunciation lexicons by looking at acoustic confusion matrices. If an ASR system consistently misrecognizes an out-of-vocabulary name as a specific sequence of known words, the phonetic properties of those errors can be reverse-engineered to deduce the true pronunciation of the name.

Instead of waiting for errors to occur, systems can also actively refine their dictionaries using acoustic data and human feedback loops:

  • Acoustic Refinement: Antoine Laurent et al. introduced an iterative method that generates multiple phonetic hypotheses for a proper noun and uses acoustic alignment scores from real speech data to filter out incorrect transcriptions.
  • Crowd-Supervised Self-Healing: To scale this process without relying on expert linguists, Ian McGraw et al. built an automated pipeline that outsources audio collection and transcription micro-tasks to non-expert crowd workers, allowing spoken language systems to automatically correct missing or broken name pronunciations.
  • Hybrid Learning: Attapol T. Rutherford et al. demonstrated that combining automated G2P algorithms directly with crowd-sourced audio data yields transcriptions for complex named entities that can even outperform expert-generated dictionary entries.

Strategy 4: Flexible Sequence Modeling

At the algorithmic level, proper names have also forced a shift away from rigid sequence alignments. Traditional G2P systems often assumed a strict one-to-one mapping between letters and phonemes. This assumption falls apart with names containing complex character clusters that collapse into a single sound, or single letters that imply multiple sounds.

To provide the necessary flexibility, Sittichai Jiampojamarn et al. introduced a many-to-many alignment framework using Hidden Markov Models. This allows variable-length chunks of letters to map directly to variable-length chunks of phonemes, naturally capturing the irregular structural patterns common in proper nouns.

Rigid 1-to-1 Alignment (Fails on irregular names):
Graphemes:  s  c  h  m  i  d  t
            │  │  │  │  │  │  │
Phonemes:   ʃ  -  -  m  ɪ  t  -

Flexible Many-to-Many Alignment (Captures structural chunks):
Graphemes:  [s c h]  [m]  [i]  [d t]
               │      │    │     │
Phonemes:     [ʃ]    [m]  [ɪ]   [t]

As deep learning advanced, neural sequence-to-sequence (Seq2Seq) models encountered similar difficulties with rare names and loanwords due to data sparsity. Benjamin Milde et al. addressed this by applying multitask learning to Seq2Seq architectures. By training a single model across multiple distinct phonetic lexicons and alphabets simultaneously, the network is forced to learn robust, generalized representations. This shared structural knowledge prevents the model from overfitting to standard vocabulary rules, significantly reducing error rates on highly irregular names and abbreviations.

Key Foundational Work

Evaluating the pronunciation of proper names by four French grapheme-to-phoneme converters (2005) — contributed 1 concept elements to this direction.
Comparative objective and subjective evaluation of three data-driven techniques for proper name pronunciation (2005) — contributed 1 concept elements to this direction.
Morphological Analysis for a German Text-to-Speech System (1986) — 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_grapheme_to_phoneme_conversion_challenge_3f5123cf,
  title = {grapheme-to-phoneme conversion challenges (grapheme-to-phoneme conversion for proper names)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grapheme-to-phoneme-conversion-challenges-grapheme-to-phoneme-conversion-for/rdr_3f5123cf2f9a4902ba67abcb266243d3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_grapheme_to_phoneme_conversion_challenge_3f5123cf_v1,
  title = {grapheme-to-phoneme conversion challenges (grapheme-to-phoneme conversion for proper names)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grapheme-to-phoneme-conversion-challenges-grapheme-to-phoneme-conversion-for/rdr_3f5123cf2f9a4902ba67abcb266243d3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake