AFX Chain Recognition in Audio Processing

• AFX Chain Recognition is an approach that infers the ordered sequence of audio effects from wet signals, recovering both effect types and their processing order.
• It comprises two main formulations: blind parameter inference for a fixed topology and order-aware classification using hyperbolic geometry.
• Empirical results indicate that hyperbolic mapping yields slight but consistent gains in ordering accuracy over Euclidean methods while ensuring functional sound reproduction.

AFX Chain Recognition denotes the inference of an ordered audio-effects processing chain from an observed wet signal. In the present literature, two formulations are central. One is blind parameter inference for a predefined chain topology, where the goal is to estimate parameters of a known analysis chain so that it reproduces a target transformation. The other is order-aware chain classification, where the task is to predict both effect identities and their order directly from wet audio. These formulations differ in supervision, identifiability, and what counts as a correct recovery: the former can yield a functionally equivalent chain realization rather than the true original chain, whereas the latter treats the entire ordered chain as the class label (Peladeau et al., 2023, Wada et al., 28 Jul 2025).

1. Task definition and formal scope

In the explicit order-aware formulation, an AFX chain is an ordered sequence of distinct effects, including the empty chain. If there are $F$ available effect types and a maximum chain length $L$, the label space is defined as

$|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$

The input is a wet signal $\mathbf{s}$, and the output is one class $a \in \mathcal{A}$, that is, the full ordered chain rather than an unordered effect set. This formulation is motivated by the fact that AFX order is non-commutative: a chain such as distortion $\rightarrow$ chorus is not equivalent to chorus $\rightarrow$ distortion, especially when non-linear or time-varying processors are involved (Wada et al., 28 Jul 2025).

This task is materially different from presence-only AFX detection. Presence-only prediction collapses distinct permutations into the same label and therefore discards the very structure that determines the final sound. It is also different from parameter estimation in a fixed chain, because the latter assumes that the effect order and topology are already known. AFX Chain Recognition, in its strongest sense, therefore concerns joint recovery of effect type and order from the processed signal.

2. Fixed-template blind estimation of known chains

Blind Estimation of Audio Effects (BE-AFX) addresses a narrower inverse problem. The setup uses a dry input $x$, a processed target $y$, unknown synthesis effects $\{e^s\}$, an analysis network $L$0, and an analysis chain $L$1. In the chain experiment, the topology is fixed to

$L$2

The analysis network predicts normalized parameters $L$3, maps them into actual parameters $L$4, and applies the analysis chain to the dry signal to obtain $L$5. What is recovered is therefore not an arbitrary chain graph, but a parameterization of a predefined chain template (Peladeau et al., 2023).

The architecture is auto-encoder-like but with a structured decoder: the encoder consumes the wet audio $L$6, produces a time-invariant embedding, and an MLP with layers of sizes $L$7, $L$8, $L$9, and $|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$0 predicts the chain parameters. Three encoders are compared: MEE, TE, and TFE. The chain blocks are differentiable DSP modules or neural approximations. Equalization uses either a 5-band parametric EQ or a 10-band graphic EQ; compression uses a simplified DSP compressor, an NP compressor, or a Hybrid NP compressor; clipping uses a parametric clipper, a Taylor clipper, or a Chebyshev clipper. Training compares two objectives. The proposed audio-based objective minimizes

$|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$1

whereas the baseline minimizes normalized parameter MSE,

$|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$2

The significance of this formulation lies in its identifiability assumptions. It does not infer whether an effect is present or absent in a variable graph, which effect classes were used among many possibilities, or the ordering of effects. When $|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$3, the model may learn parameters for an analysis chain that yields perceptually or spectrally similar audio without matching the true synthesis parameters. This supports a functional interpretation of chain recovery rather than strict causal identification.

3. Order-aware classification in hyperbolic space

A more direct treatment of AFX Chain Recognition casts the problem as multi-class classification over ordered chains. The proposed method uses raw wet waveform as input, MERT as the frontend representation model, a three-block projection MLP, an exponential map into the Poincaré ball, and hyperbolic multinomial logistic regression (MLR) as classifier. The hyperbolic formulation is motivated by two properties of the chain space: its tree-structured combinatorics and its non-commutative ordering (Wada et al., 28 Jul 2025).

The geometric model uses the $|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$4-dimensional Poincaré ball of curvature $|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$5,

$|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$6

with exponential map at the origin

$|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$7

Order sensitivity is further aligned with the non-commutativity of Möbius addition,

$|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$8

The Euclidean embedding is first computed and then mapped into hyperbolic space; the Euclidean baseline replaces this mapping with the identity and uses standard Euclidean MLR.

The experimental setup instantiates the task with three effects—delay, chorus, and distortion—each appearing at most once, with maximum chain length three. Under $|\mathcal{A}|=\sum_{l=0}^{L}\frac{F!}{(F-l)!}.$9 and $\mathbf{s}$0, the chain space contains $\mathbf{s}$1 classes, including the empty chain. Effects are applied using Pedalboard with randomized parameters over realistic ranges, and the dataset is generated from Dataset 4 of the IDMT-SMT-GUITAR corpus. In this configuration, AFX Chain Recognition becomes explicit wet-only prediction of the entire ordered chain.

4. Empirical behavior and comparative results

The two formulations produce different empirical signatures. Fixed-template BE-AFX is best read as a study of transform recovery under known topology, whereas order-aware hyperbolic classification measures direct chain recognition accuracy over ordered classes. The reported results are summarized below (Peladeau et al., 2023, Wada et al., 28 Jul 2025).

Formulation	Reported result	Main implication
Fixed chain $\mathbf{s}$2	Audio training: waveform $\mathbf{s}$3, audio loss $\mathbf{s}$4, parameter MSE $\mathbf{s}$5; parameter training: waveform $\mathbf{s}$6, audio loss $\mathbf{s}$7, parameter MSE $\mathbf{s}$8	Audio-loss training better recovers the audio transform, parameter-loss training better estimates parameters
Order-aware hyperbolic classification	Best hyperbolic: $\mathbf{s}$9, $a \in \mathcal{A}$0, Macro $a \in \mathcal{A}$1, Micro $a \in \mathcal{A}$2	Suitable curvature yields modest but consistent gains over Euclidean classification
Euclidean baseline for chain classification	Best Euclidean: $a \in \mathcal{A}$3, Macro $a \in \mathcal{A}$4, Micro $a \in \mathcal{A}$5	Ordered chain classification is feasible without hyperbolic geometry but slightly weaker
Presence-only evaluation	Euclidean: macro/micro $a \in \mathcal{A}$6; Hyperbolic: $a \in \mathcal{A}$7	Effect presence is easy; the hard part is order

These results establish a consistent distinction between effect presence and effect order. In the order-aware study, both Euclidean and hyperbolic models are nearly perfect when order is ignored, so the improvement from hyperbolic geometry arises primarily from better modeling of ordering rather than better AFX-type identification. In the fixed-chain study, by contrast, parameter supervision can outperform audio supervision on single effects such as compression and clipping, yet at the whole-chain level the audio-loss objective better reconstructs the transformation. This suggests that chain interactions can amplify parameter mismatches even when per-parameter MSE is low.

Further analysis shows that longer chains are harder to classify, that chains containing both chorus and delay are particularly confusable, and that confusion across different chain lengths remains relatively low. A plausible implication is that chain length and effect presence are learned earlier than exact permutation structure.

5. Limitations, assumptions, and common misconceptions

A central misconception is to treat all AFX-related inverse problems as equivalent. They are not. Fixed-template BE-AFX assumes a known chain structure and known order; it does not discover variable topology, infer stage presence or absence, or identify arbitrary plugin classes. Its reconstruction also still uses the dry signal $a \in \mathcal{A}$8, so it is not a fully wet-only recovery problem in deployment terms (Peladeau et al., 2023).

The explicit classification formulation is closer to unrestricted chain recognition, but it remains controlled. The reported study is synthetic, uses guitar only, limits the vocabulary to chorus, delay, and distortion, permits each effect at most once, and evaluates all possible chains within that small closed space. It does not address unseen effect categories, repeated effects, open-set recognition, unseen chain orders, or longer pedalboard-like graphs. The gains over Euclidean space are consistent but modest, and curvature must be tuned as a hyperparameter over $a \in \mathcal{A}$9, with $\rightarrow$0 giving the most reliable gains (Wada et al., 28 Jul 2025).

Another misconception is that parameter accuracy is a sufficient proxy for chain recovery. The fixed-chain results show otherwise: a model can produce better parameter MSE and worse audio-transform estimation. This is especially important when synthesis and analysis effects differ, when parameters interact nonlinearly, or when multiple parameter settings yield similar perceptual outcomes. Conversely, successful full-chain classification does not imply plugin-level identifiability or parameter recovery.

6. Terminological scope and adjacent chain-based paradigms

The word chain is used differently across adjacent literatures. In Malaysian car-plate recognition, Freeman Chain Codes represent isolated character boundaries as directional sequences; there, chain refers to contour traversal rather than ordered effect composition (Jusoh et al., 2011). In Kaldi-style ASR, chain refers to LF-MMI sequence-discriminative acoustic modeling with reduced frame rate and special HMM topology; c-DcAE, pc-DcAE, and hc-DcAE augment such chain models with latent code partitions and decoder losses, but they do not address audio-effects order inference (Lee et al., 2022). In full-text ASR post-processing, Chain of Correction denotes segment-by-segment multi-turn LLM correction with full-text context, again a sequential refinement mechanism rather than AFX-chain recovery (Tang et al., 2 Apr 2025).

Other nearby uses are conceptually informative. FakeChain treats deepfakes as ordered manipulation histories and shows that detectors often rely on the final manipulation type rather than cumulative traces, with F1-score dropping by up to $\rightarrow$1 when the final manipulation differs from training distribution (Heo et al., 20 Sep 2025). A plausible implication is that history-sensitive recognition problems can degenerate into last-stage recognition when models are trained only on final outputs. ExpLLM, by contrast, uses chain of thought for facial expression recognition, where chain denotes structured reasoning from action units to expression labels rather than signal-processing composition (Lan et al., 2024).

Within this broader landscape, AFX Chain Recognition is most precisely defined as order-sensitive inference over audio-effects compositions. Its strongest current form is wet-only ordered chain classification; its most mature inverse-control form is blind parameter estimation for a known chain template. The field’s open technical frontier remains the gap between these two regimes: variable topology, broader effect vocabularies, repeated effects, real-world wet-only deployment, and recovery of ordered processing history beyond closed synthetic chain spaces.