svara

svara (Sanskrit: voice/tone/musical note) — Formant and vocal synthesis crate for the AGNOS project.

Provides a complete formant-based vocal synthesis pipeline: glottal source generation, vocal tract modeling, phoneme-level synthesis, prosodic control, and sequenced speech rendering.

Features

Glottal source: Rosenberg pulse model with jitter, shimmer, breathiness, spectral tilt
Formant filtering: Cascade biquad resonators with Peterson & Barney (1952) vowel targets
Vocal tract: Formant filters + nasal coupling (anti-formant) + lip radiation
Phoneme inventory: ~44 phonemes covering English and major languages (vowels, plosives, fricatives, nasals, approximants, diphthongs)
Prosody: F0 contours, intonation patterns (declarative, interrogative, continuation, exclamatory), stress
Voice profiles: Male/female/child presets with formant scaling, vibrato, builder pattern
Phoneme sequences: Coarticulated rendering with crossfade transitions
Serde support: All types serialize/deserialize

Usage

use svara::prelude::*;

// Synthesize a single vowel
let voice = VoiceProfile::new_male();
let samples = synthesize_phoneme(
    &Phoneme::VowelA,
    &voice,
    44100.0,
    0.5,
).expect("synthesis should succeed");

// Build a phoneme sequence
let mut seq = PhonemeSequence::new();
seq.push(PhonemeEvent::new(Phoneme::VowelA, 0.15, Stress::Primary));
seq.push(PhonemeEvent::new(Phoneme::NasalN, 0.08, Stress::Unstressed));
seq.push(PhonemeEvent::new(Phoneme::VowelI, 0.15, Stress::Secondary));
let audio = seq.render(&voice, 44100.0).expect("render should succeed");

Feature Flags

Flag	Default	Description
`naad-backend`	Yes	Use naad crate for oscillators, filters, and noise generation
`logging`	No	Enable tracing-subscriber for structured log output

Without naad-backend, svara uses internal minimal implementations (PRNG noise, biquad filters) and compiles standalone.

Architecture

GlottalSource ──> VocalTract ──> Output
                    │
          ┌─────────┼─────────┐
          │         │         │
    FormantFilter  Nasal   Lip Radiation
    (5 biquads)   Coupling  (HPF)

Higher-level constructs:

Phoneme: Articulatory targets + synthesis strategy per class
Prosody: F0 contour + stress + intonation
VoiceProfile: Speaker parameterization
PhonemeSequence: Coarticulated multi-phoneme rendering

Consumers

dhvani — AGNOS voice/sound subsystem
vansh — AGNOS voice AI shell (TTS/STT)

License

GPL-3.0