AI model replicates Tuvan throat singing with 38% better accuracy

Tuvan sygyt singing—a biphonic technique combining a low drone with a piercing overtone in the 1–3 kHz region—has long resisted accurate digital reproduction. Standard articulatory models fail because they lack the narrow, precisely tuned resonances required. Now, researchers from an international team (Cámara et al.) have built a differentiable Kelly–Lochbaum waveguide augmented with a sublingual second source, cubic B-spline tract parameterization, and spatially varying learnable damping. The entire pipeline is optimized end-to-end via gradient descent directly from audio, allowing the model to learn the intricate vocal tract configurations that produce sygyt's characteristic merged formant structure.

Tested on 20 segments from two independent datasets (5 singers, 10 pitches), the model cut log-spectral distance by 30–38% relative to a standard articulatory baseline, with the largest gains in the overtone region. Cepstral-envelope analysis confirmed more accurate recovery of the merged formants. Notably, the model also outperformed a DDSP harmonic-plus-noise baseline that had direct per-harmonic spectral control. This suggests that building in explicit acoustic structure (a waveguide with articulatory parameters) provides a powerful inductive bias for overtone-singing copy-synthesis. The work was accepted to DAFx 2026 and is available on arXiv.