Semantic Audio Effects Mapping for Reverberation via Rule-Anchored Fusion

Abstract

Mapping free-form natural language to editable reverberation parameters remains unresolved for music education and studio practice. Cloud large-language-model pipelines are constrained by network dependence, per-call cost, data-egress risk, and second-scale round-trip latency, whereas purely embedding-based retrieval mappers—although deployable offline—are magnetically attracted to discrete anchors and emit step-like parameter trajectories when prompts vary along a continuous semantic gradient. We present a two-stage decoupled, fully offline system. The semantic-translation stage maps text to a seven-dimensional normalized parameter vector and an inspectable JSON record through rule-anchored fusion of an acoustic rule branch, a k-nearest-neighbour retrieval branch, and a residual multilayer perceptron branch; the digital-signal-processing (DSP) stage renders wet audio from dry input under a JSON-to-plugin parameter contract. Evaluation adopts a dual-axis protocol on the ISO65 anchor catalogue and the normative engineering benchmark GT v2, reporting held-out parameter mean absolute error (MAE) separately from trajectory continuity along six pre-registered real-language gradients. Under the rulefix baseline (seeds 42–44), configurations differ within one order of magnitude on the accuracy axis: held-out typical MAE (n=12) is 0.1027\pm0.0016 for k-NN, 0.0937\pm0.0045 for MLP, 0.0949\pm0.0034 for StaticEqual and 0.1004\pm0.0037 for Hybrid; the retrieval baseline remains strongly competitive and no accuracy advantage is claimed for the fused system. On the continuity axis, the fused system reduces focus-dimension trajectory irregularity relative to k-NN on all six paths, with cross-path relative improvements of 57.6\%\pm7.1\% in the variance of successive differences and 30.9\%\pm5.4\% in total variation (6/6 same direction; path-level one-sided Wilcoxon and exact sign-flip permutation p=0.0156, the minimum attainable value under 2^{6}=64 assignments). An eight-configuration extended ablation shows that removing the rule branch collapses equal-weight k-NN+MLP towards the k-NN smoothness regime (variance and total variation degrade by 43.3\% and 20.8\% relative to StaticEqual; 6/6; p=0.0156). Mechanistic diagnostics show that the rule branch behaves as a piecewise-constant, quantized-step function (adjacent-step freeze rate 35.4\%; non-zero jump-magnitude standard deviation 0.085 versus 0.210 for k-NN). Continuity gains are therefore attributed to a standard bias–variance tradeoff in which a low-variance discrete anchor is averaged with higher-variance continuous branches, not to confidence-based dynamic weighting and not to any unfalsifiable architectural emergence. The DSP stage exhibits monotonic objective responses for the injected seven dimensions, and a measured local deployment profile gives 18.71\pm3.42\,\mathrm{ms} end-to-end CPU inference, 707\,\mathrm{MB} resident memory and 88\,\mathrm{MB} on-disk dependencies. The results position rule-anchored fusion as a deployable, inspectable route to smoother semantic-to-parameter trajectories without collapsing accuracy and continuity into a single score. Keywords: semantic audio effects; reverberation parameter mapping; rule-anchored fusion; trajectory continuity; dual-axis evaluation; bias–variance tradeoff; deployability-first