voice-g2p

English grapheme-to-phoneme conversion for Kokoro TTS. A Rust port of misaki's English G2P pipeline.

Install

[dependencies]
voice-g2p = "0.1"

Usage

// Convert text to Kokoro-compatible phonemes
let phonemes = voice_g2p::english_to_phonemes("Hello world")?;
// => "həlˈO wˈɜɹld"

// For long text, split into chunks that fit the model's 510-token limit
let chunks = voice_g2p::text_to_phoneme_chunks("A very long paragraph...")?;
for chunk in &chunks {
    // Each chunk is ≤500 phoneme characters
    println!("{chunk}");
}

Custom configuration

If uv or espeak-ng aren't on your $PATH:

let config = voice_g2p::G2PConfig {
    uv_path: "/opt/homebrew/bin/uv".into(),
    espeak_path: "/opt/homebrew/bin/espeak-ng".into(),
};
let g2p = voice_g2p::G2P::with_config(config);
let phonemes = g2p.convert("Hello world")?;

What's inside

• Dictionary lookup — 90k gold + 93k silver pronunciation entries embedded at compile time
• Morphological decomposition — -s, -ed, -ing suffix rules with voicing logic
• Number handling — cardinals, ordinals, years, currency, phone numbers
• POS tagging — optional spaCy subprocess (via uv run) for context-dependent pronunciation
• Fallback — espeak-ng per-word for unknown words

Optional dependencies

• espeak-ng — fallback pronunciation for words not in the dictionary (brew install espeak-ng)
• uv — runs spaCy for POS-based disambiguation (e.g. "read" as past vs. present tense)

Both are optional. Without them, the pipeline still works using dictionary lookup alone.

License

MIT