DDSP-Violin: Unsupervised Recovery of Bowing Controls Through Interpretable Source Parameters

Duarte J. DDSP-Violin: Unsupervised Recovery of Bowing Controls Through Interpretable Source Parameters. In: AES Show 2026 Nashville. Audio Engineering Society; 2026.pp. 36:1-10. Available from: https://aes.org/publications/elibrary-page/?id=23468

@inproceedings{Duarte2026_23468,
  author = {Duarte, Joao},
  title = {{DDSP-Violin: Unsupervised Recovery of Bowing Controls Through Interpretable Source Parameters}},
  booktitle = {AES Show 2026 Nashville},
  pages = {36:1-10},
  url = {https://aes.org/publications/elibrary-page/?id=23468},
  year = {2026},
  month = oct,
  publisher = {Audio Engineering Society},
  address = {Gaylord Opryland Resort}
}

TY  - CPAPER
TI  - DDSP-Violin: Unsupervised Recovery of Bowing Controls Through Interpretable Source Parameters
AU  - Duarte, Joao
T2  - AES Show 2026 Nashville
SP  - 1
EP  - 10
M1  - Paper 36
PY  - 2026
DA  - 2026/10/30
UR  - https://aes.org/publications/elibrary-page/?id=23468
PB  - Audio Engineering Society
CY  - Gaylord Opryland Resort
LA  - en
AB  - Neural models are now the state of the art for audio synthesis, yet often offer limited interpretable control over sound production. To improve control and explainability, the Differentiable Digital Signal Processing (DDSP) framework is a promising approach for sound synthesis of acoustic musical instruments. This paper extends and evaluates DDSP-Violin, a differentiable bowed-string synthesiser whose decoder outputs a small set of parameters that target known acoustic behaviours of the instrument. To test whether these parameters capture the intended effects, the model is trained and evaluated on a controlled synthetic dataset from a mechanistic physical model with known bowing conditions. Trained unsupervised on reconstruction, the model learns a brightness parameter that correlates with bow force and speed and a bow-position parameter that recovers the node order. Ablations show that this recovery depends on an informed, constrained model design, suggesting a route towards physically meaningful, low-dimensional control in neural synthesisers.
ER  -