Automated Parameter Sensitivity Ranking in Audio Synthesis

Ogulcan Kibar; Baris Ozcan

cs.SDOct 28, 2026

Automated Parameter Sensitivity Ranking in Audio Synthesis

Authors: Ogulcan Kibar, Baris Ozcan

Organizations: Graduate School & Burhaniye Vocational School, Bahçeşehir University & Balikesir University, İstanbul, Turkey & Balıkesir, Türkiye · Computer Engineering Department, Bahçeşehir University, İstanbul, Turkey

Abstract

Software synthesizers expose dozens of coupled controls, many of them redundant or perceptually inert at typical operating points. This paper presents a two-stage, measure-and-validate framework that ranks parameters by their influence on the audio output and then checks that ranking against an independent, variance-based experiment. Every rendered clip is normalized to -23 LUFS and described by five spectral fingerprints, so that gain shifts cannot masquerade as timbral change. The Sweep Test scores each parameter by spectral deviation over a 101-step sweep; the Sobol Test scores it by summed per-cell variance over a 128-point Sobol excitation. Across five architecturally distinct synthesizers (TorchSynth, Pyo, Pedalboard, Csound, and FluidSynth) spanning 48 to 76 parameters, the two tests agree at Spearman ρ= 0.82-0.96 (p-values \le 7.2 \times 10^{-15}) on the four engines that render deterministically, and at ρ= 0.32 on the one with process-level render noise--a diagnostic of that engine rather than of the method. The strong correlations survive reducing the Sobol budget from 128 to 32 points, and the complete analysis costs a one-time offline budget of about 66,000 renders.

Figures & tables

Panel of Figure 1 — Fig. 1: Overview of the framework.

Fig. 2: Per-cell variance of the Sobol Test: each row is one of the $K$ conditioned clips, variance is taken down each cell $c$ , and the sum over cells gives $V_{i}$ .

Fig. 3: Cumulative share of the total Sweep Test score in rank order. Dashed line: $90\%$ of the total; markers: where each engine reaches it.

Rank	TorchSynth	Pyo	Pedalboard	Csound	FluidSynth
1	env-1 release	amp release	filter mode	filter type	mod-env sustain †
2	env-2 release	filter type	filter cutoff	master tune	expression
3	LFO-1-env release	filter resonance	delay time	filter cutoff	stereo balance
4	LFO-2-env release	delay mix	delay mix	osc-1 detune	pan
5	osc-2 tuning	delay time	master tune	delay mix	pitch bend
$S_{\mathrm{rank\,1}}$	0.92	0.95	1.00	1.00	1.00 †

TABLE I: Top-5 parameters per backend by

S_{i}

. † Render-noise artefact (Section V-D ).

by Sweep Test $S_{i}$			by Sobol Test $V_{i}$
Rank	Parameter	$S_{i}$	Rank	Parameter	$V_{i}$
1.5	env-1 release	0.92	1.0	osc-2 tuning	491 809
1.5	env-2 release	0.92	2.5	env-1 release	436 892
3.5	LFO-1 level env rel.	0.72	2.5	env-2 release	436 892
3.5	LFO-2 level env rel.	0.72	4.0	osc-2 mod depth	436 844
5.0	osc-2 tuning	0.68	5.5	LFO-1 level env rel.	307 692

TABLE II: TorchSynth top-15 by the Sweep Test score

S_{i}

(left) and by the Sobol Test variance

V_{i}

(right).

Function	TorchSynth	Pyo	Pedalboard	Csound	FluidSynth
Amp-env release	0.92 (1)	0.95 (1)	0.29 (7)	0.00 (46)	0.00 (22)
Amp-env attack	0.06 (21)	0.03 (9)	0.02 (45)	0.22 (8)	0.00 (47)
Filter cutoff	—	0.03 (8)	0.42 (2)	0.37 (3)	0.03 (6)
Filter mode/type	—	0.60 (2)	1.00 (1)	1.00 (1)	—
Delay mix	—	0.29 (4)	0.31 (4)	0.26 (5)	—
Master tune	0.68 (5)	0.00 (22)	0.31 (5)	0.37 (2)	0.00 (43)

TABLE III: Functionally equivalent parameters across engines:

S_{i}

with rank in parentheses.

Backend	$N$	$\rho(S,V)$	$p$ -value	$\rho(S,V^{\mathrm{agg}})$	Agreement
TorchSynth	76	$+0.961$	$6.6\times 10^{-43}$	$+0.948$	strong
Pedalboard	56	$+0.923$	$4.7\times 10^{-24}$	$+0.951$	strong
Csound	52	$+0.874$	$2.9\times 10^{-17}$	$+0.959$	strong
Pyo	56	$+0.823$	$7.2\times 10^{-15}$	$+0.820$	strong
FluidSynth	48	$+0.319$	$2.7\times 10^{-2}$	$+0.303$	weak

TABLE IV: Spearman

\rho

between the

S

and

V

rankings at

K=128

, for the headline log-STFT score and the five-map aggregate.

Engine	Parameter	Identifier	$S_{i}$	$S$ -rank	$V$ -rank
Pedalboard	filter mode	flt.mode	1.00	1	1
Csound	filter type	flt.type	1.00	1	1
TorchSynth	envelope-1 release	adsr1.release	0.92	1	3
TorchSynth	osc-2 tuning	vco2.tuning	0.68	5	1
Pyo	amp-env release	amp.release	0.95	1	2
Pyo	delay time	del.time	0.22	5	42

TABLE V: Ten parameters of special interest, with

S_{i}

and both rank positions.

	Leave-one-out (worst of 5)			Scaling of $S_{i}$
Backend	$S$	$V^{\mathrm{agg}}$	valid.	pctl.	rank
TorchSynth	$+0.985$	$+0.964$	$+0.919$	$+0.994$	$+0.987$
Pedalboard	$+0.991$	$+0.988$	$+0.932$	$+0.991$	$+0.983$
Csound	$+0.985$	$+0.974$	$+0.944$	$+0.975$	$+0.983$
Pyo	$+0.869$	$+0.938$	$+0.711$	$+0.812$	$+0.821$
FluidSynth	$+0.967$	$+0.999$	$+0.287$	$+0.870$	$+0.889$

TABLE VI: Robustness ablations.

Fig. 4: TorchSynth — rank-1 parameter: envelope-1 release ( adsr1.release , $S_{i}{=}0.92$ ).

Audio Generation

Audio generation creates speech, music, and sound from text, media, symbolic controls, or reference audio, with growing focus on controllability and realism.

See all 134 papers

Synth-JEPA: Joint Embedding Prediction for Renderer-Free Synthesizer Parameter Search
Sep 25, 2026Ben Hayes, Haokun Tian, Stefan LattnerAudio Representation Learning Joint-Embedding Predictive Architecture
Sony Computer Science Laboratories, Paris, France · Queen Mary University of London, UK
INSTRUMENTAL: Automatic Synthesizer Parameter Recovery from Audio via Evolutionary Optimization
Mar 16, 2026Philipp BogdanParameter Estimation Evolutionary Optimization
FM Synthesizer Audio-Parameter Shared Embeddings
Aug 18, 2026David Braun, Adam FinkelsteinMusic Information Retrieval Graph Neural Networks
Dept. of Computer Science Princeton University Princeton, NJ, USA

Sep 25, 2026J/K move · Enter open · S save

Synth-JEPA: Joint Embedding Prediction for Renderer-Free Synthesizer Parameter Search

Ben Hayes, Haokun Tian, Stefan Lattner

Sony Computer Science Laboratories, Paris, France · Queen Mary University of London, UK

Audio Representation Learning Joint-Embedding Predictive Architecture Parameter Estimation Latent Space Optimization

Open paper AskPDF arXiv

Sound matching can be formulated as optimizing synthesizer parameters against an audio-domain objective. However, objectives derived from generic audio representations are often difficult to optimize, while direct search requires rendering every candidate. We introduce Synth-JEPA, which learns mutually predictive audio and parameter representations from paired synthesizer data. At inference, candidate parameters are scored directly in this learned space, yielding a renderer-free objective whose audio geometry is shaped by parameter correspondences rather than generic audio similarity. We evaluate Synth-JEPA on Surge XT using held-out synthesizer sounds and out-of-domain NSynth and FSD50K targets, against inverse models, direct search, and learned proxy objectives. Synth-JEPA outperforms all baselines in-domain and remains competitive out-of-domain. Its matching quality continues to improve with additional test-time search, allowing compute to be traded for match quality. In pairwise listening tests, listeners preferred Synth-JEPA in 85% of trials overall. Together, these results show that an audio representation with a parameter-induced geometry allows synthesizer sound matching to be approached as an effective renderer-free search problem.

Abstract

Audio Generation

Explore similar work

Synth-JEPA: Joint Embedding Prediction for Renderer-Free Synthesizer Parameter Search

INSTRUMENTAL: Automatic Synthesizer Parameter Recovery from Audio via Evolutionary Optimization

FM Synthesizer Audio-Parameter Shared Embeddings