The word "read" is a linguistic trap for a computer. In the sentence "I will read the book," it rhymes with bead; in "I have read the book," it rhymes with bed. These words—homographs—share a spelling but diverge in sound and meaning. For a Text-to-Speech (TTS) system to sound natural, it must resolve this ambiguity before a single phoneme is synthesized.
The core challenge in Grapheme-to-Phoneme (G2P) conversion is that the correct pronunciation often lives outside the word itself. While many words follow standard phonetic rules, homographs require a broader view of the sentence to unlock their correct sound.
The Limitation of Isolated Word Models
Traditional G2P systems often process text one word at a time. As Martin Jansche argues in Re-Engineering Letter-to-Sound Rules, models that look at words in isolation are fundamentally incapable of resolving pronunciation changes that depend on neighboring words or grammatical roles. When a system only sees the string "read," it has no way to decide between its two phonetic realizations.
This "word-level" blindness leads to systematic errors. Wasala et al. note that without higher-level linguistic information, systems cannot reliably distinguish between words that are orthographically identical but phonetically distinct. Similarly, Zhu et al. point out that failing to consider surrounding text prevents models from resolving ambiguities in languages where identical spellings represent different sounds based on syntactic cues.
Scaffolding with Syntax and Semantics
To bridge this gap, researchers have historically turned to explicit linguistic features. The intuition is that if a model knows a word's Part-of-Speech (POS), it can narrow down the pronunciation. For instance, the word "object" is stressed on the first syllable as a noun ("an OB-ject") but on the second as a verb ("to ob-JECT").
Braga et al. demonstrate that combining POS parsing with semantic analysis can significantly improve the naturalness of synthesized speech by resolving these heterophonic homographs. El-Kahky et al. further show that specifically targeting noun-verb ambiguity can slash pronunciation errors, as these are among the most common points of failure for automated readers.
Beyond simple tags, the structural relationship between words provides even deeper cues. Soong et al. found that incorporating features from a syntactic parse tree—such as the "syntactic distance" between words—helps the model understand the phrasing and rhythm required for complex sentences. This structural awareness prevents the system from mispronouncing words that change sound based on their function within a phrase.
The Shift to Sentence-Level Transformers
Modern approaches have moved away from manual rule-based systems toward end-to-end neural architectures that ingest entire sentences at once. By using Transformers, these models can use "attention" to look at the entire context simultaneously, effectively performing disambiguation and phonetic conversion in a single step.
Řezáčková et al. introduced T5G2P, which treats G2P as a text-to-text task using the T5 transformer. By training on full sentences, the model learns to pick up on the subtle contextual clues—like the presence of an auxiliary verb—that signal a homograph's correct reading. Rahmati et al. applied a similar sentence-level approach to Persian, a language where missing short vowels make homograph disambiguation particularly critical.
The power of this sentence-level view is its robustness. Zhao et al. show that by leveraging the surrounding sentence context, models can even "auto-correct" pronunciations when faced with typos or OCR errors. Instead of relying on a single, potentially misspelled word, the model uses the collective evidence of the entire phrase to infer the intended sound. As Jůzová et al. suggest, while phrase-level context is a massive leap forward, the ultimate goal remains a unified system that seamlessly integrates these broad linguistic patterns to ensure every "read" is heard exactly as the author intended.
Key Foundational Work
A Phonological Processor for Italian (1983) — contributed 3 concept elements to this direction.
A Sentence Analysis Method for a Japanese Book Reading Machine for the Blind (1986) — contributed 1 concept elements to this direction.
An Expert System For The Production Of Phoneme Strings From Unmarked English Text Using Machine-Induced Rules (1983) — contributed 1 concept elements to this direction.