shabda

shabda (Sanskrit: word / sound) — Grapheme-to-phoneme (G2P) conversion for Rust.

The bridge between text and vocal synthesis. Converts English and Spanish text to phoneme sequences ready for svara's synthesis engine. Uses shabdakosh's 10,000+ entry pronunciation dictionary with intelligent rule-based fallback.

Features

• Multi-language G2P: English (General American) and Spanish (Castilian)
• 10,000+ word dictionary: Via shabdakosh — CMUdict-derived, O(1) lookup, variant pronunciations
• Rule-based fallback: Context-sensitive letter-to-sound rules for unknown words
• Silent letter handling: knight, gnome, write, psychology, lamb, etc.
• Morphological awareness: -tion/-sion suffixes, -ed endings (/t/ vs /d/ vs /id/), un-/re-/dis- prefixes
• Magic-e and r-colored vowels: make→/meik/, car→/kar/, bird→/berd/
• Syllabification: Maximal Onset Principle with sonority constraints
• Syllable-weight stress: Heavy penult rule for polysyllabic words
• Prosody control: Phrase pauses, emphasis markers (CAPS, *asterisks*), speaking rate (WPM), timing profiles
• SSML support: <break>, <emphasis>, <prosody> tags via convert_ssml()
• Streaming API: Word-by-word callback via convert_streaming()
• Accuracy: Abbreviation expansion (Dr.→doctor), acronym handling (NASA vs FBI), foreign word detection, heteronym disambiguation
• Number expansion: 42→"forty two", 3.14→"three point one four"
• speak() method: Text to audio samples in one call (G2P + svara rendering)
• varna integration: Phoneme inventory validation and language detection (optional)
• Serde support: All types serialize/deserialize
• no_std compatible: Works with alloc, no standard library required

Quick Start

use shabda::prelude::*;

// Convert text to phonemes
let g2p = G2PEngine::new(Language::English);
let events = g2p.convert("hello world").unwrap();

// With options: emphasis + slow rate
let opts = ConvertOptions::new()
    .with_emphasis(true)
    .with_speaking_rate(100.0);
let events = g2p.convert_with("HELLO world", &opts).unwrap();

// SSML input
let events = g2p.convert_ssml(
    r#"Hello <break time="300ms"/> <emphasis level="strong">world</emphasis>"#
).unwrap();

// Spanish
let g2p_es = G2PEngine::new(Language::Spanish);
let events = g2p_es.convert("hola mundo").unwrap();

// Direct to audio
let voice = svara::voice::VoiceProfile::new_male();
let samples = g2p.speak("hello world", &voice, 44100.0).unwrap();

Architecture

Input text
    |
    v
Abbreviation Expansion (Dr. → doctor)
    |
    v
Acronym Handling (FBI → f b i, NASA → nasa)
    |
    v
Number Expansion (42 → "forty two")
    |
    v
Normalizer (lowercase, punctuation → phrase markers, emphasis markers)
    |
    v
G2P Engine
    |-- Heteronym check (read/read, live/live — context-based)
    |-- Dictionary lookup (shabdakosh, 10K+ entries)
    |-- Foreign word detection (strip diacritics, retry)
    |-- Rule-based fallback (English or Spanish rules)
    v
Syllabifier (Maximal Onset Principle)
    |
    v
Prosody Mapper (stress, emphasis, rate, timing)
    |
    v
Vec<PhonemeEvent> (ready for svara)

Feature Flags

Consumers

• dhvani — AGNOS audio engine (text-to-speech pipeline)
• vansh — Voice AI shell
• Any application needing text-to-speech with svara

License

GPL-3.0-only