svara

svara (Sanskrit: स्वर — voice / tone / musical note) — Formant and vocal synthesis for Rust.

Complete formant-based vocal synthesis pipeline: dual glottal source models (Rosenberg B + LF), SOA-vectorized formant filtering, 48 phonemes, prosodic control, look-ahead coarticulation, and spectral analysis. Built on hisab for math and naad for DSP primitives.

Features

• Dual glottal models: Rosenberg B polynomial + Liljencrants-Fant (LF) with Rd voice quality parameter
• SOA formant filter: Structure-of-arrays biquad bank (MAX_FORMANTS=8) with compiler auto-vectorization — 2x faster than scalar
• 48 phonemes: 15 vowels, 5 diphthongs, 6 plosives, 9 fricatives, 3 nasals, 4 approximants, 2 affricates, glottal stop, tap/flap, silence
• Hillenbrand formant data: Per-vowel frequencies and bandwidths from Hillenbrand et al. (1995)
• Vocal tract: Parallel formant bank + nasal coupling (place-dependent) + subglottal resonance + lip radiation + source-filter interaction + DC blocking + gain normalization
• Prosody: Monotone cubic f0 contours, 4 intonation patterns, stress, Catmull-Rom interpolation
• Coarticulation: Look-ahead onset, sigmoid crossfades, per-phoneme resistance coefficients (Recasens DAC), F2 locus equations
• Voice profiles: Male/female/child presets with f0-dependent bandwidth scaling, vibrato, builder pattern
• Spectral analysis: FFT-based spectrum, formant estimation, band energy, compensated RMS
• Performance: ~1,000x real-time, f64 biquad coefficients, no_std compatible, all types Send + Sync

Quick Start

use svara::prelude::*;

let voice = VoiceProfile::new_male();
let samples = synthesize_phoneme(&Phoneme::VowelA, &voice, 44100.0, 0.5).unwrap();

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).unwrap();

Feature Flags

Architecture

GlottalSource (Rosenberg/LF) → VocalTract → Output
                                  │
                    ┌─────────────┼──────────────┐
                    │             │              │
              FormantFilter   Nasal         Lip Radiation
              (SOA biquad     Coupling       + Subglottal
               bank, 8-wide)  (place-dep)    + Interaction

Consumers

• dhvani — AGNOS audio engine
• vansh — Voice AI shell (TTS/STT)
• prani — Creature vocal synthesis (depends on svara)

License

GPL-3.0-only