We present a differentiable, time-domain implementation of an extended Karplus-Strong Algorithm (KSA) that can be optimised end-to-end as a Differentiable DSP (DDSP) paradigm. We show its ability for instrument modelling by reconstructing an acoustic guitar dataset, where our proposed Gradient Descent (GD) baseline outperforms a Genetic Algorithm (GA).
Figure 1: Proposed architecture, where coloured blocks are optimised by gradient descent. Lagrange is short for Lagrange interpolation, while g stands for gain stages.
| Baseline | \(\mathcal{L}_{\mathrm{MSS}}\) | MFCC | Encodec | MS-CLAP |
|---|---|---|---|---|
| Gradient Descent | 0.79 ± 0.11 | 2.32 ± 1.50 | 19.89 | 111.30 |
| Genetic Algorithm | 1.43 ± 0.31 | 4.10 ± 1.75 | 216.93 | 346.93 |
| Ground Truth | Gradient Descent | Genetic Algorithm |
|---|
| \(\downarrow\) Oct. Vel. \(\rightarrow\) | 25 | 50 | 75 | 100 | 127 | mean |
|---|---|---|---|---|---|---|
| E2–E3 | +0.27 | +0.23 | +0.33 | +0.30 | +0.36 | +0.30 |
| E3–E4 | -0.02 | +0.07 | -0.02 | +0.07 | +0.01 | +0.02 |
| E4–E5 | -0.18 | -0.10 | -0.05 | -0.13 | -0.08 | -0.11 |
| E5–E6 | -0.19 | -0.26 | -0.27 | -0.18 | -0.11 | -0.20 |
| mean | -0.03 | -0.01 | +0.00 | +0.02 | +0.05 | +0.00 |
Click any cell in the table above to explore audio examples for that specific combination of pitch and dynamics.
| Ground Truth | Gradient Descent |
|---|
If you use this work, please cite:
Tablas de Paula, P., Marttila, D. and Reiss, J.D., 2025 December. Differentiable Karplus-Strong. In Digital music research network one-day workshop (DMRN+20).
@inproceedings{tablas2025diffks,
title = {Differentiable Karplus-Strong},
author = {Tablas de Paula, Pablo and Marttila, David and Reiss, Joshua D.},
booktitle = {Digital Music Research Network One-day Workshop (DMRN+20)},
year = {2025},
month = dec
}