Digital audio workstations expose rich effect chains, yet a semantic gap remains between perceptual user intent and low-level signal-processing parameters. We study retrieval-grounded audio effect control, where the output is an editable plugin configuration rather than a finalized waveform. Our focus is Texture Resonance Retrieval (TRR), an audio representation built from Gram matrices of projected mid-level Wav2Vec2 activations. This design preserves texture-relevant co-activation structure. We evaluate TRR on a guitar-effects benchmark with 1,063 candidate presets and 204 queries. The evaluation follows Protocol-A, a cross-validation scheme that prevents train-test leakage. We compare TRR against CLAP and internal retrieval baselines (Wav2Vec-RAG, Text-RAG, FeatureNN-RAG), using min-max normalized metrics grounded in physical DSP parameter ranges. Ablation studies validate TRR's core design choices: projection dimensionality, layer selection, and projection type. A near-duplicate sensitivity analysis confirms that results are robust to trivial knowledge-base matches. TRR achieves the lowest normalized parameter error among evaluated methods. A multiple-stimulus listening study with 26 participants provides complementary perceptual evidence. We interpret these results as benchmark evidence that texture-aware retrieval is useful for editable audio effect control, while broader personalization and real-audio robustness claims remain outside the verified evidence presented here.
Figures & tables
Work
Input
Output
Grounding
Editability
SAFE [ 5 ]
Text
Parameters
None
Yes
DDSP [ 3 ]
Audio
Parameters
None
Yes
LLM2Fx [ 37 ]
Text
Parameters
None
Yes
AudioLDM [ 1 ]
Text
Waveform
Pretrained
No
HM-RAG [ 38 ]
Text+Doc
Text
Retrieval
N/A
TRR-based retrieval control (Ours)
Text+Audio
Parameters
Dual-Modal Retrieval
Yes
TABLE I: Positioning Against Prior Work
Fig. 1: System overview of the retrieval-grounded audio effect control pipeline. The real-time DSP engine uses a six-module serial effect chain (as shown in the framework), where each module can be independently bypassed. Retrieved parameters are validated against per-module physical bounds (Eq. 1 ) before being applied.
Fig. 2: Gram matrix heatmaps for representative query presets. Each matrix shows co-activation structure across projected channels, revealing distinct texture patterns for different effect styles.
Component
Median (ms)
p95 (ms)
Text scoring
3.027
3.586
Audio scoring (TRR cosine)
1.369
1.564
Fusion
0.110
0.146
Projection
0.031
0.039
End-to-end (local pipeline)
4.538
5.269
TABLE II: Per-query local retrieval-path latency breakdown (cache-only). We report median and p95 over 633 runs from the original 211-query profiling pool ( 211×3 repeats).
TABLE III: Protocol-A: Direct retrieval comparison on the resolved-audio-grouped split ( Ntest=204 , Nkb=1063 ). Lower is better for L2; higher is better for all other metrics. TRR achieves the best mean performance in every reported metric; the CLAP row uses the 201 shared queries with cached CLAP embeddings.
Scenario
Reference branch
Fusion (Ours)
Interpretation
Vague Text
Text-only: 12.6517
0.2970
fallback to audio
Noisy Audio
TRR-only: 33.8549
1.5148
fallback to text
Conflict
TRR-only: 0.2970
4.5259
failure mode
TABLE IV: Selected Protocol-C boundary cases ( N=211 ). The current fusion heuristic mainly acts as a fallback rule: it collapses onto the intact modality under single-modality degradation, but degrades substantially under explicit modality conflict.
Fig. 5: t-SNE comparison of TRR (Gram matrix) and Wav2Vec2 mean-pooled embeddings on the full dataset, colored by style label. TRR exhibits tighter within-style clustering, supporting its utility as a texture-aware retrieval prior.