{"id":"4a8b36df-ca7b-4590-b535-32b8130cc07e","arxiv_id":"2505.01998","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A product-oriented preprint claims that combining nonlinear acoustic equations with reinforcement learning improves far-field speech processing, but its benchmark results come from separate commercial systems and do not validate the proposed framework.","lead":"This paper sketches a framework that combines nonlinear acoustic wave equations (Westervelt/KZK) with reinforcement learning for robot speech processing, then reports benchmark numbers from the authors' commercial speech products. A generalist might read it to see how a company frames physics-based audio and RL as a path to better noise suppression, but the experiments do not actually test the framework.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4 benchmarks validate standalone Azero products, not the Section 2 Westervelt/KZK+RL framework; Section 6.1 concedes real-world validation is future work, so the claimed superiority is unsupported.","rationale":"The reader identified the load-bearing assumption as the claimed connection between Section 2's nonlinear acoustic/RL framework and Section 4's benchmark results. My reading confirms this: the paper's experimental section evaluates commercial product components (AzeroVEP, AzeroASR, AzeroTTS, AzeroGPT) without demonstrating that these components implement the Westervelt/KZK equations or the PPO controller. Section 6.1 explicitly says real-world validation is future work, which is an in-scope limitation statement that supports the reader's concern. The strongest claim in the abstract and conclusion is therefore not supported by the presented evidence. I also noted a concrete technical error in the KZK equation's nonlinear term, which reinforces the impression that the theoretical framework is not actually exercised by the experiments. Since the verdict is already REJECT and this concern does not change that assessment, I recommend UNCHANGED. Agreement with the reader is 'agree' because the same weakest assumption was identified.","tokens_in":18718,"tokens_out":4481,"duration_ms":44089,"concrete_test":"Inspect the repository at github.com/soundai2016/nonlinear-acoustic-rl-hri and trace the evaluation pipeline for one benchmark, e.g., Fleurs ASR from Table 4. Determine (a) whether any code path solves Eq. (2) or (5) (or a KZK/Westervelt PDE), and (b) whether PPO from Eq. (6) updates any parameters during inference. If neither is present, rerun the same benchmark with the nonlinear/PDE component ablated, or replaced by a linear propagation model, and compare WER and MOS-LQO. If results are statistically unchanged, the Section 4 numbers validate only the Azero products, not the proposed framework.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that embedding nonlinear Westervelt/KZK acoustics in an RL control loop (Section 2) yields superior far-field localization, weak-signal detection, and multilingual ASR. For that claim to hold, the Section 4 results must be produced by, or at least depend on, the Section 2 framework. The paper never shows this. Section 3 presents AzeroVEP, AzeroASR, AzeroTTS, and AzeroGPT as engineering products, and Section 4 reports their benchmark numbers (Tables 1-10). Nowhere is it stated that AzeroVEP solves Eq. (2) or (5), that the PPO objective in Eq. (6) updates absorption or beamforming parameters during these runs, or that the ASR/TTS/GPT results change if the nonlinear acoustic model is removed. Section 6.1 explicitly concedes: \"the current work is based on evaluations using existing benchmark datasets, and the true potential of this approach has yet to be fully explored in real-world environments.\" Thus the experiments cannot distinguish the proposed hybrid from ordinary deep-learning baselines; the claimed advantage over \"traditional linear methods and purely data-driven baselines\" is not established by any presented comparison. A secondary correctness issue compounds the gap: Eq. (3)-(5) write the KZK nonlinear term as d^3(p^2)/dtau^3, whereas the standard KZK/Westervelt paraxial form uses d^2(p^2)/dtau^2 with prefactor beta/(2 rho0 c^3). As printed, the physical equations are not the equations claimed, further undermining the link between theory and the reported product results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid framework that integrates nonlinear acoustic wave equations (Westervelt and KZK) with reinforcement learning for far-field human-robot interaction. Section 2 presents the physical equations and a PPO-based adaptive control loop; Section 3 describes a set of commercial products (AzeroVEP, AzeroASR, AzeroTTS, AzeroGPT); Section 4 reports benchmark results for those products; Section 5 surveys applications; Section 6 concludes that the hybrid surpasses classical linear and purely data-driven baselines. The paper also states in Section 6.1 that only existing benchmark datasets were used and that real-world validation remains future work.","tokens_in":19145,"tokens_out":3225,"duration_ms":31808,"significance":"If the central claim were established, embedding nonlinear acoustic propagation models inside an RL-controlled front-end could be a meaningful direction for robot audition and far-field speech processing. The paper articulates that ambition and provides standard textbook equations, but the manuscript's actual strengths are limited: there is no parameter-free derivation, no machine-checked proof, and no experimental protocol that links the Section 2 theory to the Section 4 results. The benchmark tables are extensive but, as written, cannot be used to validate the proposed framework. The explicit concession in Section 6.1 that only existing benchmark datasets were used further weakens the central claim of real-world superiority.","major_comments":[{"comment":"The experimental section is disconnected from the proposed framework. The benchmark results in Tables 1-10 evaluate AzeroVEP, AzeroASR, AzeroTTS, and AzeroGPT as standalone products, but the paper never states that these systems implement the Westervelt/KZK equations (Eqs. 2-5) or the PPO objective in Eq. (6). Section 6.1 explicitly concedes that 'the current work is based on evaluations using existing benchmark datasets,' so the claimed superiority over traditional linear methods and purely data-driven baselines is not supported by any presented comparison. This is the load-bearing gap of the paper.","section":"§4, §6.1"},{"comment":"The KZK equation is written with a nonlinear term ∂³(p²)/∂τ³, but the canonical KZK/Westervelt paraxial form uses ∂²(p²)/∂τ² with prefactor β/(2ρ0c³); only the absorption term, δ/(2c³)∂³p/∂τ³, is third-order in τ. As printed, the physical equations are not the standard equations claimed. Since the theoretical foundation depends on these equations, this error undermines the paper's claim of a physically informed framework.","section":"§2.1.2, Eqs. (3)-(5)"},{"comment":"Headline performance improvements are asserted without experimental detail: localization error reduction from 15° to 3° and a fivefold speedup (Section 3.2), and a 58 dB voice-clarity improvement at 120 dB noise (Section 3.4). No measurement protocol, baseline definition, error bars, or statistical comparison is provided for any of these numbers. These claims are central to the paper's stated contributions and cannot be evaluated or reproduced.","section":"§3.2, §3.4"}],"minor_comments":[{"comment":"The text refers to 'Table 1' and 'Table 2' for the LibriSpeech results, but the corresponding tables are numbered Table 2 and Table 3. Please correct the cross-references.","section":"§4.2"},{"comment":"The table title misspells 'Fleurs' as 'Fluers', and the French column header is truncated to 'F'. Please fix the typo and formatting.","section":"Table 6"},{"comment":"The likelihood ratio rt(θ) and the advantage estimator Ât are used in the PPO objective but never defined. Please define rt(θ) = πθ(at|st)/πθ_old(at|st) and specify how the advantage is estimated.","section":"§2.2.2, Eq. (6)"},{"comment":"The concession that only existing benchmark datasets were used contradicts the abstract's claim of validation in 'demanding real-world scenarios.' Please reconcile these statements.","section":"§6.1"},{"comment":"Figures 6-11 are referenced in the text but are not present in the provided manuscript. Please ensure they are included in the final PDF.","section":"Figures"},{"comment":"The GitHub repository link is given, but the repository contents and reproduction instructions are not described. Please state which results can be reproduced and how.","section":"Code/Data Availability"}],"recommendation":"reject","confidential_remarks":"The manuscript reads primarily as a product showcase for SoundAI's Azero line, with the nonlinear-acoustic/RL framework serving as an asserted umbrella rather than as a demonstrated technical contribution. The absence of any ablation or controlled study connecting Eqs. (2)-(6) to the benchmark numbers is a central validity issue, not a presentation issue. If the authors can produce experiments in which the Westervelt/KZK module and the RL control loop are actually present and ablated, a revised submission could be reconsidered; but as is, the central claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a company product whitepaper wrapped in a physics paper. The Section 2 framework (Westervelt/KZK + PPO) never connects to the Section 4 benchmarks, and the paper's own Section 6.1 admits real-world validation is still to come. The headline claim that the hybrid outperforms linear and purely data-driven baselines is unsupported.\n\nTo give credit: the high-level idea of using RL to adapt parameters in nonlinear acoustic models is a reasonable direction, and the paper does collect a broad set of product metrics across noise suppression, ASR, TTS, and LLM benchmarks. If the linked GitHub repo actually ships reproducible code for those numbers, that's an asset. That's about where the credit ends.\n\nSoft spots, in order of severity. First, the disconnect between theory and experiment is load-bearing. Nowhere do the authors state that AzeroVEP or AzeroASR actually solves Eq. (2) or (5), or that the PPO objective in Eq. (6) updates any acoustic parameter during the reported runs. The benchmarks are for standalone commercial systems, so they can't validate the claimed framework. Second, the paper itself concedes this: Section 6.1 says the current work is based on existing benchmark datasets and the true potential has yet to be explored in real-world environments. That directly contradicts the abstract's claim of superior performance in demanding real-world scenarios. Third, there is a concrete technical error: Eq. (3)-(5) have ∂³(p²)/∂τ³ in the nonlinear term, but the standard KZK equation uses ∂²(p²)/∂τ². As printed, these are not the equations the paper says they are. Fourth, the headline numbers (12 dB SNR gain, 58 dB clarity improvement, localization 15° to 3°, 5x speedup) appear with no experimental procedure, error bars, or baseline definitions. The PPO section is just one equation; no state/action/reward details, no training curves, no evidence the RL loop was ever run.\n\nWho is this for? Not a research audience. It might be a useful overview for someone evaluating SoundAI's products, or a strawman for a blog post about the gap between paper claims and product reality. As a scientific submission, it doesn't hold together.\n\nRecommendation: desk reject. The idea is worth a real paper someday, but this manuscript would need a complete rewrite—controlled ablations, actual connection of the products to the framework, correct equations, and error bars—before it deserves referee time.","headline":"A product whitepaper dressed as a research paper: the nonlinear-acoustic+RL framework is never actually connected to the reported benchmarks, and the paper's own conclusion concedes real-world validation is still future work.","tokens_in":19651,"tokens_out":3667,"would_cite":false,"duration_ms":34619,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that nonlinear acoustic equations (Westervelt/KZK) tuned in real time by a reinforcement-learning controller can beat both linear methods and pure deep learning for noisy human-robot interaction.","keywords":["nonlinear acoustics","Westervelt equation","KZK equation","reinforcement learning","human-robot interaction","far-field speech recognition","noise suppression","multilingual speech recognition"],"falsifier":"Run the Section 4 benchmark suite with the same Azero pipeline but with the Westervelt/KZK nonlinear terms zeroed and the PPO controller frozen; if MOS-LQO, WER, and localization error stay essentially unchanged, the experiments are not testing the proposed framework's contribution.","tokens_in":18573,"feed_emoji":"🎙️","tokens_out":10335,"duration_ms":93953,"temperature":0.7,"pith_summary":"The paper proposes a robot-audition framework that couples the Westervelt and KZK nonlinear acoustic equations with a reinforcement-learning controller, using the nonlinear terms to model harmonic generation and shock formation while the RL agent retunes absorption, beamforming, and filter parameters in real time. The intended contribution is a physics-informed alternative to purely data-driven pipelines that stays interpretable yet adapts to nonstationary noise and reverberation. The authors claim this hybrid strategy surpasses linear methods and deep-learning baselines in far-field source localization, weak-signal detection, and multilingual speech recognition, citing gains like 12 dB SNR improvement, localization error cut from 15 degrees to 3 degrees, and low real-time factors. They also position the same architecture for AI hearing aids, smart speakers, in-vehicle systems, and brain-machine interfaces. The paper is explicit in Section 6.1 that evaluations so far use existing benchmark datasets, with real-world deployment left as next steps.","feed_headline":"Nonlinear acoustics plus RL outdoes linear and deep baselines","feed_subtitle":"The hybrid Westervelt/KZK + PPO loop is claimed to lift SNR by 12 dB and cut localization error from 15° to 3°.","key_machinery":"The load-bearing object is the Westervelt equation $\\partial^2 p/\\partial t^2 - c^2\\nabla^2 p = \\alpha\\,\\partial^2(p^2)/\\partial t^2$, whose quadratic term acts as a source for second-harmonic generation and shock formation, together with the paraxial KZK extension that adds diffraction and thermoviscous absorption. Around these equations the paper wraps a PPO-driven control loop whose actions are incremental adjustments to the model parameters ($\\alpha$, $\\beta_{NL}$, $\\delta$, beamforming weights) and whose reward balances recognition accuracy, latency, and energy use. The nonlinear terms supply physical structure the RL agent can exploit, while the RL agent supplies the real-time adaptation that the bare equations lack.","core_discovery":"On the paper's own account, the central discovery claim is that embedding physically informed nonlinear wave equations in a reinforcement-learning control loop yields a self-tuning acoustic front end whose performance exceeds linear processing and pure deep learning in demanding acoustic environments. The proposed mechanism is a feedback loop: the Westervelt and KZK equations provide the physical model of harmonic generation, shock formation, and diffraction, while a PPO agent continuously adjusts propagation coefficients, filter gains, and beamforming weights to match changing conditions. The reported evidence comes from the Azero product family: AzeroVEP reaches MOS-LQO 4.29 versus 2.8 for RNNoise at 20 dB babble noise, AzeroASR reaches WER 3.86 percent and 5.12 percent on the Fleurs Chinese and English sets, and localization error drops from about 15 degrees to 3 degrees with a fivefold throughput gain. Section 6.1 concedes these results come from existing benchmark datasets and that real-world validation is future work.","pith_inferences":["The paper reports product-level benchmarks for AzeroVEP, AzeroASR, AzeroTTS, and AzeroGPT without showing that those products execute the Section 2 equations and PPO loop; the convincing test would be an ablation that disables the nonlinear terms and the RL controller without touching the rest of the pipeline.","If the nonlinear-RL loop is genuinely responsible for the gains, the method should transfer most strongly to regimes where nonlinearity is unavoidable, such as high-intensity focused sound, shock-forming fields, and extreme reverberation, and should be compared there against linear beamforming plus a standard neural enhancer.","A clean experimental extension would measure how many RL interactions the controller needs to re-adapt when the room impulse response or noise type changes; that number would tell whether the 'real-time adaptation' claim holds on edge hardware."],"forward_implications":["If the framework works as claimed, far-field speech recognition in industrial and traffic noise can hold above 96 percent accuracy, with a 12 dB SNR improvement in 100 dB environments.","Nonlinear terms give the front end a physical account of harmonic generation and shock formation, which should make it more reliable at high sound pressure levels and in strongly reverberant enclosures than linear models.","The RL loop replaces manual hand-tuning: the same pipeline can be deployed in factories, cars, auditoriums, and homes without per-environment parameter engineering.","On-device deployment becomes realistic because reported real-time factors stay at or below roughly 0.1, and the voice-cloning and AzeroGPT modules inherit the noise-robust front end."],"supporting_citations":[{"why":"It supplies the nonlinear acoustics background and the Westervelt/KZK equations used in Section 2.","marker":"[2]"},{"why":"It provides the deep-reinforcement-learning methods, including PPO-style policy optimization, used as the adaptive controller.","marker":"[6]"},{"why":"It supplies the supervised speech-separation baseline against which the noise-suppression improvements are measured.","marker":"[15]"},{"why":"It is the RNNoise hybrid DSP/deep-learning baseline directly compared with AzeroVEP in Figure 6.","marker":"[16]"},{"why":"It supplies the FireRedASR benchmark numbers that AzeroASR is compared against on Chinese speech.","marker":"[18]"},{"why":"It supplies the Kimi-Audio LLM-based ASR baseline numbers on Fleurs and AISHELL.","marker":"[19]"},{"why":"It supplies the Whisper ASR baseline used for the multilingual Fleurs word-error-rate comparison.","marker":"[22]"}],"fun_headline_variants":["Physics-informed RL tunes acoustic waves for robots","Nonlinear wave equations plus RL beat deep and linear baselines","PPO-driven acoustic model sharpens robot speech and localization","Hybrid Westervelt-KZK and RL improve real-world HRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the assumption that the benchmark results in Section 4 come from systems that actually use the proposed nonlinear acoustic plus reinforcement-learning framework; the paper never states that they do, and Section 6.1 says only existing benchmark datasets were used.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed RL tunes acoustic waves for robots","Nonlinear wave equations plus RL beat deep and linear baselines","PPO-driven acoustic model sharpens robot speech and localization","Hybrid Westervelt-KZK and RL improve real-world HRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1704,"prompt_tokens":910,"completion_tokens":794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":727}},"tokens_in":526,"tokens_out":794,"duration_ms":9105,"temperature":1.0,"reasoning_tokens":727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:04:05.960439+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Section 4 benchmark suite with the same Azero pipeline but with the Westervelt/KZK nonlinear terms zeroed and the PPO controller frozen; if MOS-LQO, WER, and localization error stay essentially unchanged, the experiments are not testing the proposed framework's contribution.","supporting_citations":[{"cited_title":"A survey on deep reinforcement learning for audio-based applications,","cited_arxiv_id":null,"evidence_quote":"It provides the deep-reinforcement-learning methods, including PPO-style policy optimization, used as the adaptive controller."},{"cited_title":"Supervisedspeechseparationbasedondeeplearning: Anoverview,","cited_arxiv_id":null,"evidence_quote":"It supplies the supervised speech-separation baseline against which the noise-suppression improvements are measured."},{"cited_title":"A hybrid dsp/deep learning approach to real-time full-band speech enhancement,","cited_arxiv_id":null,"evidence_quote":"It is the RNNoise hybrid DSP/deep-learning baseline directly compared with AzeroVEP in Figure 6."},{"cited_title":"Robustspeechrecognition via large-scale weak supervision,","cited_arxiv_id":null,"evidence_quote":"It supplies the Whisper ASR baseline used for the multilingual Fleurs word-error-rate comparison."}],"review_version":1}