{"id":"6b2bd142-5424-49c7-959f-1c3620b89dac","arxiv_id":"2505.04082","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Stretched tanh activations reduce measured aliasing in neural amp models by up to about 27 percent relative to the standard tanh, with a small increase in waveform-matching error.","lead":"The paper shows that using smoother nonlinear curves inside neural guitar-amp models cuts digital aliasing, and introduces a new metric, the Aliasing-to-Signal Ratio, to measure it. A generalist might care because this is a zero-compute fix for a common artifact in neural audio emulation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ASR is asserted to measure aliasing but is never validated, and the sine-test amplitude and transient handling are unreported; the 27% headline depends on these unspecified choices.","rationale":"In good faith, the paper's core idea is plausible and partly supported: the relationship between smooth nonlinearities and reduced high-order harmonic energy is theoretically expected, the study uses many seeds, and the spectral plots for selected CustomTanh variants illustrate the effect. The reader's conditional verdict is appropriate. However, the single most load-bearing assumption is not merely that ASR is perceptually relevant, but that ASR accurately measures aliasing at all in the reported experiments. This requires the analyzed output to be exactly periodic and free of transient/non-aliasing energy, and it requires a well-specified input amplitude. The paper neither validates ASR against a known aliasing ground truth nor reports these experimental parameters, and no code or data are released for independent verification. This is a measurement-validity concern rather than a disagreement with the qualitative trend. If the proposed synthetic check shows ASR tracks true aliased energy across amplitudes and transient handling, the central claim stands and the paper's main limitation becomes a need for listening tests and code release; if it does not, the numerical headline is unsupported. Since the issue is concrete and addressable, the reader's CONDITIONAL verdict remains the right call, with no change recommended.","tokens_in":18394,"tokens_out":14244,"duration_ms":154935,"concrete_test":"Implement the ASR measurement on a controlled memoryless tanh nonlinearity at fs=48,017 with a 1249 Hz sine. Sweep input amplitudes from -40 to 0 dBFS, and compute ASR (a) on the full one-second output and (b) after discarding the first 2045 samples (the receptive field) and taking exactly N=48,017 steady-state samples. Compare against a ground-truth alias ratio obtained by 8x oversampled simulation of the same nonlinearity. If ASR deviates from ground truth by more than 3 dB, or if the CustomTanh alpha=1 versus alpha=1.8 ordering changes with amplitude or transient handling, the headline 27% reduction is not robust as reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that smoother activations reduce aliasing, with CustomTanh alpha=1.8 giving a 27% ASR reduction—rests on the ASR metric, yet ASR is treated as ground truth without validation. ASR defines EA = EY - EH, so every bin not at a low-order harmonic is counted as aliasing. This is exactly correct only if the analyzed output is a periodic steady-state single-sine response. The paper states this condition ('provided that any transient response is discarded') but reports no input amplitude, no transient-discard length, and no initial-state handling. The 18-layer dilated TCN has a receptive field of 2045 samples (kernel 3, dilation pattern repeated 1..256); if the one-second test window includes the onset transient, energy leaks across all bins and is misclassified as aliasing. Because the same ASR ranks all 124 configurations, a systematic leakage or amplitude-dependent bias would directly change the reported ranking and the 27% figure. The abstract's claim that ASR 'quantitatively assesses the level of aliasing with high accuracy' is thus an assertion, not a demonstrated property; no synthetic ground-truth comparison or ablation is provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether smoother activation functions reduce aliasing artifacts in neural amplifier models built as temporal convolutional networks (TCNs). The authors introduce a new metric, the Aliasing-to-Signal Ratio (ASR), defined as the ratio of non-harmonic spectral energy to harmonic energy in the response to a periodic sine test. They train 12,400 models across 124 activation configurations (62 functions, gated and non-gated, with 100 seeds) and then 20,000 additional models to sweep the stretch/oscillation parameter for CustomTanh and Snake. They report that smoother activations, especially CustomTanh with larger stretch factors, reduce ASR while keeping the Error-to-Signal Ratio (ESR) modest, with a highlighted 27 percent ASR reduction for CustomTanh_1.8 relative to the baseline, accompanied by a modest ESR increase. The paper also includes spectrum and waveform analyses for three CustomTanh variants.","tokens_in":18614,"tokens_out":5583,"duration_ms":54215,"significance":"If the ASR metric is valid and the reported effect is robust, the paper provides a practically useful, architecture-free way to reduce aliasing in a widely used neural amp modeling framework: simply choose a smoother activation function. The experimental scale is a genuine strength: 100 deterministic seeds per configuration, 12,400 plus 20,000 trained models, and alpha-sweep experiments give the main trend empirical support rather than making it an artifact of a single run. The ASR definition is mathematically clean under the stated periodicity and coprimality assumptions, and no parameters are fitted to force the smoothness conclusion. However, the significance of the central quantitative claims currently rests on an unvalidated metric and a sine-test protocol whose key details are unreported; these issues must be addressed for the paper's headline numbers to be credible.","major_comments":[{"comment":"The ASR definition EA = EY - EH counts every DFT bin not at a low-order harmonic as aliasing, which is exact only for the periodic steady-state response to a single sine. The paper states in Section 3.2 that this requires 'any transient response is discarded,' but Section 4.4 reports no input amplitude, no transient-discard length, and no initial-state handling. The 18-layer TCN has a receptive field of 2045 samples (kernel 3 with the dilation pattern in Section 2.4), so a one-second test window that includes the onset will leak energy across all bins and misclassify that leakage as aliasing. Because the same ASR ranks all configurations, this missing detail directly affects the reported ranking and the 27 percent figure; the setup must be specified or the claim limited.","section":"Section 3.2 and Section 4.4"},{"comment":"The abstract states that ASR 'quantitatively assesses the level of aliasing with high accuracy,' but no validation of the metric is provided. There is no comparison with a synthetic signal whose aliasing content is known exactly, no comparison with an established perceptual or objective aliasing measure, and no listening test (deferred in Section 5). Since the paper's central conclusion is an ASR-based ranking of 124 activation configurations, this is a load-bearing gap. A concrete validation, such as adding known aliasing components to a clean periodic signal and checking whether ASR tracks the added energy monotonically, or correlating ASR with listening-test ratings, would support the claim.","section":"Section 3.2 and abstract"},{"comment":"The headline 27 percent reduction (ASR 0.001582 for False_CustomTanh_1.8 versus 0.002173 for False_CustomTanh_1) is measured with a single sine at 1249 Hz and no reported amplitude. Nonlinear aliasing depends strongly on drive level and input frequency; a single point does not establish that the improvement holds across the operating range of a guitar amplifier model. The authors should report an amplitude sweep and at least several fundamental frequencies, or explicitly restrict the claim to the tested condition.","section":"Section 4.4 and Table 2"},{"comment":"The exclusion of ReLUSquared and ReLUSquaredDip from the minimum-ASR analysis because they 'failed to train effectively (ESR about 1)' is a post-hoc criterion that is not stated in the methodology and is applied only to the minimum-ASR half of Table 1. The text acknowledges this, but the minimum-ASR ranking then depends on an unprincipled filter. Either state a rule (e.g., exclude all configurations with average ESR above a threshold) before presenting the table, or report the unfiltered results so readers can see the effect.","section":"Section 4.2.3"}],"minor_comments":[{"comment":"Training hyperparameters (optimizer, learning rate, batch size, and the epoch count at early stopping) are not stated; 'replicated from [15]' is insufficient for reproducibility of a study whose quantitative claims rest on 100-seed averages.","section":"Sections 2.2 and 2.4"},{"comment":"The statement that aliasing 'remains below -60 dB' for alpha = 1.8 is based on visual inspection of the right column of Figure 6; the authors should define the quantitative criterion, such as the maximum magnitude of non-harmonic bins below 6 kHz.","section":"Section 4.4"},{"comment":"In the submitted PDF, the figure labels and legend text contain garbled font-encoding sequences (e.g., '/uni00000014...'), making the plots unreadable; the final version must render these characters properly.","section":"Figures 3-5"},{"comment":"The gated versus non-gated comparison confounds the choice of activation function with the additional sigmoid nonlinearity in the gate; the conclusion that 'gating introduces aliasing' should be qualified as applying to the gated architecture as a whole rather than to the activation alone.","section":"Section 4.1.1"},{"comment":"The 'balanced' points (alpha = 1.8 for Tanh, alpha = 2.9 for Snake) are chosen informally; the authors should state the selection rule, such as the lowest ASR subject to an ESR ceiling, to make the choices reproducible.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of DAFx, and the large-scale experiments are a genuine strength. The main obstacle to publication is that the central quantitative claims depend on an ASR metric that is asserted, not validated, and on a sine-test protocol whose amplitude and transient handling are unreported. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a useful, incremental paper, not a breakthrough. The qualitative claim that smoother nonlinearities alias less is old news in audio DSP. What's new is the systematic sweep: 124 activation configurations, 100 seeds each, plus finer alpha sweeps for Tanh and Snake, and a dedicated metric (ASR) to quantify aliasing in a sine response. The experimental effort is real and honestly reported.\n\nThe ASR definition is mathematically clean under the stated periodicity and coprimality assumptions, and the paper correctly notes that transients must be discarded. The problem is that the paper never reports the sine-test input amplitude, any transient-discard length, or initial-state handling. The 18-layer TCN has a receptive field of 2045 samples. If the one-second test window includes the onset transient, energy leaks across all bins and gets misclassified as aliasing. That could bias the ranking, not just the absolute numbers.\n\nThe 27% reduction headline (CustomTanh alpha=1.8 vs alpha=1) is also a post hoc \"balanced\" point selection, and the standard deviations for those two configurations overlap heavily. So the exact figure is shaky, even though the overall trend is consistent across many configurations.\n\nCredit where due: the ASR vs ESR trade-off analysis is useful, the finding that gated activations increase aliasing is interesting, and the adaptive-alpha suggestion is a practical takeaway. The paper also states plainly in Section 5 that listening tests are needed, which is honest.\n\nMain soft spots, in ascending order of concern: no code or data released (minor, since the training data is public NAM data but the test pipeline is part of the method); the ASR is never validated against a ground truth or perception (medium, because it's the central tool); and unreported amplitude/transient handling (medium-to-serious, because it could affect the rankings). None of these are fatal; all are addressable.\n\nWho this is for: people building neural amp models or similar nonlinear audio emulators, and researchers working on alias metrics. It deserves a serious referee.\n\nRecommendation: send to peer review. A reviewer should ask for amplitude/transient reporting, a validation of ASR on a synthetic nonlinearity where aliasing is known exactly, and ideally a small listening test. The core finding will likely survive those additions.","headline":"Solid empirical sweep showing smooth activations reduce aliasing in neural amp models; the ASR metric is promising but unvalidated, and missing test details make the 27% headline provisional.","tokens_in":19152,"tokens_out":2343,"would_cite":true,"duration_ms":24074,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Smoother, horizontally stretched tanh activations reduce aliasing in neural amplifier models while keeping modeling error modest.","keywords":["neural amp modeling","aliasing","activation functions","Aliasing-to-Signal Ratio","CustomTanh","WaveNet","error-to-signal ratio","audio effects modeling"],"falsifier":"Run a listening test comparing CustomTanh $\\alpha=1$ and $\\alpha=1.8$ on real guitar input; if listeners cannot reliably tell which model is less aliased, or if the ASR ranking reverses when the sine test is repeated at different input amplitudes, the paper's central claim fails.","tokens_in":18160,"feed_emoji":"🎸","tokens_out":7864,"duration_ms":75212,"temperature":0.7,"pith_summary":"The paper claims that the aliasing artifacts plaguing neural guitar-amplifier models are not an inevitable cost of nonlinearity: the choice of activation function can suppress them substantially. It introduces a metric, the Aliasing-to-Signal Ratio (ASR), built on a number-theoretic choice of DFT length, to measure aliasing energy separately from harmonic energy, and uses it to rank 124 activation configurations. The central finding is that smoother, ungated activations—above all the horizontally stretched tanh, $\\mathrm{CustomTanh}(x)=\\tanh(x/\\alpha)$—produce markedly less aliasing, with $\\alpha=1.8$ keeping aliasing components below $-60$ dB and cutting ASR by about 27 percent relative to the standard $\\alpha=1$ baseline while raising ESR only modestly. If this holds, neural amp models can become cleaner without architectural changes or extra computation.","feed_headline":"Stretching tanh cuts neural-amp aliasing by 27%","feed_subtitle":"Just swapping the activation function drops aliasing to -60 dB with little loss of accuracy.","key_machinery":"The load-bearing piece is the newly defined Aliasing-to-Signal Ratio (ASR), computed on the output of a single-tone sine test. The trick is to choose the DFT length $N$ and the fundamental bin $k_0$ coprime (here $N=48{,}017$, a prime, and $f_0=1249$ Hz), so that the set $\\{n\\cdot k_0 \\bmod N\\}$ fills every bin exactly once before any bin receives a second component; harmonic energy $E_H$ is read off the known harmonic bins and total spectral energy $E_Y$ from all bins, and $\\mathrm{ASR}=(E_Y-E_H)/E_H$ estimates the aliasing fraction. Because no windowing or zero padding is used, each bin samples a single frequency, making the energy subtraction exact for a periodic steady-state response. The second object is the stretched tanh, $\\mathrm{CustomTanh}(x)=\\tanh(x/\\alpha)$, whose larger $\\alpha$ flattens the nonlinearity and is shown to lower ASR monotonically while ESR passes through a minimum near $\\alpha=0.8$.","core_discovery":"The paper's central claim is that smoother activation functions consistently produce less aliasing in TCN-based neural amp models, and that this gain can be purchased with only a modest rise in waveform-matching error. In the paper's own measurements, non-gated CustomTanh with stretch factor $\\alpha=1.8$ reaches aliasing components below $-60$ dB, an ASR of $0.001582$ versus $0.002173$ for $\\alpha=1$, a 27 percent reduction, at ESR $0.016628$ versus $0.013467$. The authors also find a systematic trade-off: gated activations such as SELU achieve the lowest ESR ($0.010591$) but introduce roughly 6–9 times more aliasing than the best anti-aliasing configurations. The conclusion is that activation-function choice is a viable anti-aliasing lever that requires neither architectural change nor additional computation.","pith_inferences":["A natural extension the paper does not run: repeat the ASR measurement across input amplitudes and fundamental frequencies; the trade-off curve could flatten or invert, which would decide whether adaptive-$\\alpha$ control should be level-aware.","The paper's gating results suggest an untested lever: because gated variants alias 6–9 times more than ungated ones, the sigmoid gate path itself is a likely aliasing source, and smoothing or removing that nonlinearity could combine low ESR with low ASR.","ASR is an unweighted spectral-energy ratio; before it is used to tune production models, a perceptually weighted or music-signal version would need validation against the listening tests the paper calls for."],"forward_implications":["Swapping the standard tanh for CustomTanh with $\\alpha=1.8$ reduces aliasing components below $-60$ dB and cuts ASR by roughly 27 percent at the cost of a small ESR rise, with no change to architecture or compute.","Among 124 configurations, the ten lowest-ASR models are all ungated, while gated variants dominate the lowest-ESR list, so the ASR-ESR trade-off is systematic rather than specific to one function.","Increasing the stretch factor $\\alpha$ reduces ASR continuously in the tanh family, giving practitioners a dial between aliasing and waveform error; the paper also suggests the possibility of adapting $\\alpha$ in real time based on input pitch.","The same smoothing logic highlights Hardtanh as a near-Pareto option and motivates exploring spline-smoothed, rounded-corner variants, which the paper proposes as future work.","For the Snake family, ASR and ESR both have local minima near $\\alpha=2.9$, so Snake offers a balanced operating point distinct from the stretched-tanh family."],"supporting_citations":[{"why":"Supplies the WaveNet-based TCN architecture, the training configuration, and the standard tanh baseline that all variants are compared against.","marker":"[15]"},{"why":"Defines the original gated activation and dilated-convolution architecture underlying the model.","marker":"[12]"},{"why":"Introduces the Error-to-Signal Ratio and the pre-emphasis loss used in training and evaluation.","marker":"[13]"},{"why":"Provides the sine-wave spectral plot with circled harmonics that the paper adapts for showing aliasing.","marker":"[22]"},{"why":"Documents the aliasing problem in TCN amp models at high fundamentals that motivates the study.","marker":"[16]"},{"why":"Supplies the Snake activation family whose $\\alpha$-scaled variants are evaluated for the ASR-ESR trade-off.","marker":"[19]"}],"fun_headline_variants":["Smoother activations cut amp-model aliasing 27%","Stretched tanh slashes neural-amp aliasing, not accuracy","Activation choice trims aliasing in neural amps","Smooth curves reduce aliasing in neural amp models","Anti-aliasing via activation: no arch change, just tanh"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole ranking of activations hangs on the assumption that ASR computed from one 1249 Hz sine tone, whose input level is not reported, captures the aliasing artifacts that listeners actually hear in musical signals; the paper itself notes that listening tests are still needed.","fun_headline_variants_meta":{"raw":{"variants":["Smoother activations cut amp-model aliasing 27%","Stretched tanh slashes neural-amp aliasing, not accuracy","Activation choice trims aliasing in neural amps","Smooth curves reduce aliasing in neural amp models","Anti-aliasing via activation: no arch change, just tanh"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1205,"prompt_tokens":907,"completion_tokens":298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":211}},"tokens_in":523,"tokens_out":298,"duration_ms":3109,"temperature":1.0,"reasoning_tokens":211,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:39:21.043741+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a listening test comparing CustomTanh $\\alpha=1$ and $\\alpha=1.8$ on real guitar input; if listeners cannot reliably tell which model is less aliased, or if the ASR ranking reverses when the sine test is repeated at different input amplitudes, the paper's central claim fails.","supporting_citations":[{"cited_title":"Table 2: Comparison of Best Performing Models with Baselines","cited_arxiv_id":null,"evidence_quote":"Supplies the WaveNet-based TCN architecture, the training configuration, and the standard tanh baseline that all variants are compared against."},{"cited_title":"The Fender Bassman 5F6-A family of preamplifier circuits—a wave digital filter case study,","cited_arxiv_id":null,"evidence_quote":"Defines the original gated activation and dilated-convolution architecture underlying the model."},{"cited_title":"Wave digital mod- eling of circuits with multiple one-port nonlinearities based on Lipschitz-bounded neural networks,","cited_arxiv_id":null,"evidence_quote":"Introduces the Error-to-Signal Ratio and the pre-emphasis loss used in training and evaluation."},{"cited_title":"Real-time black-box modelling with recurrent neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the sine-wave spectral plot with circled harmonics that the paper adapts for showing aliasing."},{"cited_title":"Real-time modeling of audio distortion circuits with deep learning,","cited_arxiv_id":null,"evidence_quote":"Documents the aliasing problem in TCN amp models at high fundamentals that motivates the study."},{"cited_title":"A vacuum-tube guitar amplifier model using long/short-term memory networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Snake activation family whose $\\alpha$-scaled variants are evaluated for the ASR-ESR trade-off."}],"review_version":1}