Automated Phonetic Transcription from Orthography

24 concepts23 papers

In speech technology, the bridge between written text (graphemes) and spoken sounds (phonemes) is rarely a direct one-to-one mapping. Languages often exhibit deep orthographies where spelling does not perfectly reflect pronunciation, and even "phonetic" scripts face the challenge of out-of-vocabulary (OOV) words. Grapheme-to-phoneme (G2P) conversion is the mechanism used to predict the pronunciation of any character sequence, enabling systems to handle words they have never encountered in a formal dictionary.

The Dictionary Problem and OOV Words

Most speech recognition (ASR) and text-to-speech (TTS) systems rely on a lexicon—a static mapping of words to their phonetic realizations. However, as Hahn et al. observe, modern systems must handle a vast, evolving vocabulary that includes proper nouns, technical terms, and neologisms not present in these fixed lists.

G2P models solve this by learning the underlying statistical relationship between character sequences and phone sequences. Hahn et al. compared several methods, such as the "Sequitur" and "Phonetisaurus" tools, demonstrating that high-quality G2P conversion directly improves the accuracy of large-vocabulary continuous speech recognition. For keyword search (KWS) tasks, Liu et al. showed that G2P is essential for generating pronunciations for OOV keywords, allowing them to be integrated into search frameworks without manual transcription.

Mechanisms of Mapping

The conversion process typically involves one of three architectural approaches:

  1. Statistical and Tree-Based Models: Early robust systems used Classification and Regression Trees (CART) to predict phonemes based on local graphemic context. Udhaykumar et al. utilized these to bootstrap multilingual ASR for Indian languages, where data scarcity makes manual lexicons impractical.
  2. Finite-State Transducers (FSTs): For languages with more predictable but complex rules, FST grammars provide an efficient way to encode linguistic knowledge. Oo et al. applied this to Burmese, creating open-source grammars that normalize text and generate pronunciations for TTS.
  3. Joint Sequence Modeling: In languages like Arabic, where vowels and diacritics are often omitted in writing, the mapping is inherently ambiguous. Glass et al. demonstrated that stochastic lexicons—which produce a set of potential phonetic realizations for a single orthographic string—allow acoustic data to resolve this ambiguity during recognition.

Scaling to Low-Resource Languages

G2P conversion is particularly vital for low-resource languages where comprehensive pronunciation dictionaries do not exist. Instead of building a dictionary from scratch, researchers often leverage multilingual G2P models. Do et al. found that using a pre-trained multilingual G2P model can serve as an effective substitute for a native dictionary in TTS, yielding speech quality comparable to systems trained on ground-truth data.

Similarly, for languages like Yoruba, the use of tools like the Epitran G2P system allows for the rapid creation of phonetically-aware speech corpora by automatically converting orthographic text into phonemic representations (Gutkin et al.). This automation reduces the need for expert phonetic transcription, which Strömbergsson et al. note is a significant bottleneck in linguistic research.

Beyond Phonemes: Articulatory and Visual Features

While most G2P systems target phonemes, the mapping can be extended to more granular or different modalities:

  • Articulatory Features: Rather than mapping to discrete phone labels, systems can map graphemes to vectors of articulatory features (e.g., voiced, nasal, bilabial). Mortensen et al. introduced PanPhon to facilitate this, showing that these linguistically rich features improve performance in cross-lingual settings.
  • Visemes: In audio-visual speech synthesis, text must be mapped to mouth shapes (visemes). Weiss and Aschenberner developed viseme-based transcription for German to ensure that the visual component of a synthetic avatar remains synchronized with the generated audio.

By treating orthography as a noisy or underspecified signal of the underlying speech production process, G2P conversion provides the flexibility needed for speech systems to operate in the open-world environment of human language.

Key Foundational Work

Multilingual Speech Recognition for Information Retrieval in Indian Context (2004) — contributed 1 concept elements to this direction.
A German viseme-set for automatic transcription of input text used for audio-visual speech synthesis (2005) — 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_bdbc2189,
  title = {grapheme-to-phoneme conversion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grapheme-to-phoneme-conversion/rdr_bdbc21892b7d4a8ca0c5ffec06b8d25d},
  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_grapheme_to_phoneme_conversion_bdbc2189_v1,
  title = {grapheme-to-phoneme conversion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grapheme-to-phoneme-conversion/rdr_bdbc21892b7d4a8ca0c5ffec06b8d25d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake