G2P Conversion for OOV Words in Multilingual TTS

Grapheme-to-Phoneme (G2P) conversion translates written orthography into phonetic transcriptions, serving as a foundational step in Text-to-Speech (TTS) synthesis. In multilingual systems, Out-of-Vocabulary (OOV) words—such as proper nouns, technical jargon, neologisms, and foreign loanwords—frequently break standard dictionary lookups. Modern multilingual TTS architectures overcome this limitation by deploying neural sequence-to-sequence models, shared phonetic representations, morphological decomposition, and cross-lingual transfer techniques to predict accurate pronunciations dynamically.

Neural Sequence-to-Sequence Generalization

Traditional G2P relied heavily on static pronunciation lexicons and hand-crafted pronunciation rules, which immediately fail when encountering an OOV token. Modern pipelines replace or augment lexicons with neural sequence-to-sequence architectures, predominantly Transformer-based or recurrent neural networks (RNNs).

These models operate at the character level. By mapping grapheme sequences directly to phoneme sequences, neural models learn latent spelling-to-sound rules from vast multilingual training corpora. When an unseen word appears, the network does not require an exact match; it extrapolates phonemic output based on learned statistical patterns of surrounding graphemes, prefix-suffix structures, and phonotactic probabilities.

Unified Phonetic Spaces (IPA and X-SAMPA)

Multilingual G2P models commonly project diverse writing systems into a shared target vocabulary, such as the International Phonetic Alphabet (IPA) or X-SAMPA. Instead of maintaining isolated phoneme sets for each language, a global acoustic-phonetic inventory is used.

This unified space allows the system to utilize cross-lingual transfer. If an OOV word shares phonetic or orthographic characteristics with another language within the model's training data, the model transfers sound-mapping rules across linguistic boundaries. Low-resource languages benefit directly from high-resource languages that share linguistic roots or orthographic conventions, enabling viable zero-shot pronunciation generation.

Subword and Morphological Decomposition

In agglutinative or compounding languages (such as German, Finnish, or Turkish), OOV words are frequently new combinations of known morphemes rather than entirely novel roots. Multilingual G2P systems address this through subword tokenization techniques, such as Byte-Pair Encoding (BPE), or dedicated morphological analyzers.

By decomposing a complex OOV word into constituent stems, prefixes, and suffixes, the system converts an unknown compound into several known units. Each unit is mapped to its corresponding phonemes, and language-specific phonological rules (such as liaison, vowel harmony, or German final-obstruent devoicing) are applied at the boundaries during reconstruction.

Code-Switching and Context-Aware Language Identification

In multilingual TTS, OOV words are often loanwords or instances of code-switching (e.g., an English brand name appearing in Spanish text). Standard monolingual G2P typically mispronounces these by forcing native grapheme-to-phoneme rules onto foreign orthography.

Advanced architectures integrate Language Identification (LID) modules directly into the text front-end or condition the G2P model with language embeddings. When the model detects an OOV token from a foreign origin:

  1. It applies the phonology of the source language to generate the base pronunciation.
  2. It maps foreign phonemes to the closest equivalent sounds supported by the target speaker’s voice via phoneme substitution matrices.

Hybrid Inference Pipelines

In production systems, G2P modules typically operate as hybrid pipelines:

  1. Lexicon Lookup: The system first queries high-precision, human-curated dictionaries for known irregular pronunciations.
  2. Neural Prediction: Any unresolved tokens (OOVs) are passed to the multilingual neural G2P model.
  3. Phonotactic Constraints: The generated phonemes are checked against language-specific phonotactic constraints to avoid generating unpronounceable sound sequences before being passed to the acoustic model.