{"id":"d499fd5a-099f-4473-b4f3-feb91b5f9a0a","arxiv_id":"2505.19480","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Feeding a measured room impulse response into a neural acoustic echo canceller, especially as a synthesized echo input, improves its performance on unseen and real rooms.","lead":"This paper tests using a measured room impulse response as a side input, or prompt, to a deep acoustic echo cancellation network, and reports better echo suppression on unseen rooms and real recordings. The best fusion method denoises the RIR, convolves it with the far-end signal, and feeds that echo estimate into the model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All reported gains use clean, oracle-equivalent RIR prompts; the paper's own noisy-pulse-collection scenario is never tested, so the central generalization claim is conditional on an unvalidated denoising step.","rationale":"The paper's central claim is that RIR prompts improve AEC generalization, with fusion (d) as the strongest variant. For that claim to hold in any real deployment, the system must be able to obtain a usable RIR from a noisy pulse measurement, as the method itself describes in Section 3.1. The reported experiments, however, appear to use clean RIRs both for generating the echo and for constructing the prompt, so the prompt is a near-oracle signal. The reader's weakest assumption identifies exactly this dependence on accurate RIR measurement and denoising, and the paper's own Section 5 acknowledges the difficulty of noisy RIRs for traditional methods without providing a quantitative evaluation of its denoising stage. This is the single most load-bearing concern because it targets the boundary between the controlled experimental setting and the claimed practical applicability: if the denoiser cannot recover the RIR accurately, the prompt echo becomes a biased input and the model may be harmed rather than helped. The proposed perturbation test is concrete and feasible given the released code, and it would settle whether the reported gains are robust to realistic RIR acquisition error. Since the concern is exactly the condition already reflected in the reader's conditional verdict, no verdict adjustment is needed; the paper remains conditionally acceptable pending this test.","tokens_in":8094,"tokens_out":4680,"duration_ms":42716,"concrete_test":"Run fusion (d) on the Mismatch Synthetic and Real RIR test sets with the RIR used for the prompt perturbed to simulate realistic measurement error: (1) add ambient noise to the RIR at 10, 5, and 0 dB SNR before the denoising module, (2) shift the RIR by 1, 5, and 10 ms to simulate time alignment error, and (3) use an RIR measured from a different M-L distance or a previous room state while the echo remains generated from the true RIR. Compare PESQ, SDR, and ERLE to Table 1. If any perturbation reduces fusion (d) gains below baseline, or below half of the reported improvement, the central generalization claim is contingent on near-perfect RIR acquisition; if gains persist under all perturbations, the concern is resolved. Also report the denoiser's output RIR error or output SNR to define the operating range of the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that fusion (d) improves AEC generalization by conditioning on a measured RIR, including under the noisy pulse-collection protocol described at the end of Section 3.1. The load-bearing assumption is that the noisy RIR can be denoised well enough that the prompt echo v'(n)=RIR_d1*x(n) is a faithful estimate of the true echo. This assumption is not tested anywhere in Sections 4-5. Section 4.1 describes generating training/test RIRs with the image method and using a recorded-RIR database; no noise is added to the RIR itself. Thus RIR_n in Figure 2(d) is effectively a clean RIR, and the reported gains in Table 1 give the model near-oracle access to the exact echo path used to create the echo. Section 5 motivates the denoiser by showing a spectrum of a noisy RIR, but no quantitative denoising accuracy or any experiment with noisy or perturbed RIR prompts is reported. If the denoiser fails or the RIR is stale (people moving, gyroscope-triggered re-measurement lag), the prompt echo is biased; the network may be misled and the Table 1 gains could vanish or reverse. The paper itself flags that traditional methods fail on noisy RIR (Section 5), yet it does not demonstrate that its own method survives the same condition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes to improve the generalization of learned acoustic echo cancellation (AEC) systems by supplying the network with a room impulse response (RIR) as an auxiliary \"prompt.\" Four fusion strategies are described: (a) a learned time-frequency embedding, (b) a temporally averaged STFT frame, (c) a cropped STFT prompt, and (d) denoising/truncating the RIR and convolving it with the far-end signal to create a prompt echo v'(n). Experiments with ICCRN and MTFAA backbones on matched synthetic, mismatched synthetic, and recorded real-RIR test sets report consistent gains for fusion (d), with smaller or mixed gains for the other fusions, at a modest increase in computation and parameters.","tokens_in":8402,"tokens_out":8267,"duration_ms":62142,"significance":"If the reported results hold under realistic RIR estimation errors, the work offers a practical, low-cost way to inject environment knowledge into learned AEC systems, and the evaluation across three test sets, including real RIRs, is a strength. The release of code and audio samples is also a positive feature. However, the current evaluation is built from clean, oracle RIRs: in fusion (d) the prompt echo is constructed from the exact RIR used to generate the true echo, so the comparison is leakage-prone. The paper's central real-world premise, namely that the RIR is measured in a noisy environment, is explicitly stated but never tested. The contribution is promising but not yet established.","major_comments":[{"comment":"Fusion (d) constructs v'(n) by convolving the far-end signal with the denoised/truncated RIR. In the experiments of Section 4.1, the same image-method RIR is used to generate both the true echo v(n) and the prompt, so the model is given a near-exact copy of the echo it must cancel. The reported gains therefore do not isolate the benefit of RIR prompting. Please add ablations in which the prompt RIR is corrupted by noise, truncation, delay or gain errors, or is deliberately stale (collected at a different loudspeaker position), and report AEC metrics as a function of RIR error; also compare against a baseline that receives the oracle echo v(n) as an auxiliary input to quantify the ceiling.","section":"3.2(d), 4.1, Table 1"},{"comment":"The paper motivates the denoising module by showing a noisy RIR spectrum, but no experiment actually uses noisy or perturbed RIR prompts; all Table 1 results use effectively clean RIRs. The claim that the method works in the noisy-pulse protocol described in Section 3.1 is therefore unsupported. Please report quantitative denoising accuracy (e.g., RIR SNR or echo-path error) and downstream AEC metrics under the noisy-pulse scenario, and describe the denoising module's architecture, training target, and training data if it is learned.","section":"5 (last paragraph), Figure 4"},{"comment":"For the ICCRN baseline on the Mismatch Synthetic RIR set, fusion methods (a), (b), and (c) do not improve over the baseline (PESQ 1.94/1.95/1.96 vs 1.98; SDR 3.70/4.03/3.84 vs 4.28). Only fusion (d) improves. The general statement that RIR prompting improves generalization is therefore not supported by three of the four variants; the conclusion should be restricted to fusion (d) or accompanied by an error analysis explaining why the other fusions fail on mismatched conditions.","section":"Table 1, Mismatch Synthetic RIR"},{"comment":"No error bars, confidence intervals, or significance tests are reported, and several differences are small relative to typical per-utterance variability in AEC evaluation, e.g., Match Synthetic PESQ 2.96 to 3.03 or Real RIR ERLE 3.73 to 4.79 for ICCRN. Please report per-utterance statistics or results over multiple seeds to substantiate the abstract's claim of significant improvement.","section":"5, Table 1"}],"minor_comments":[{"comment":"The definition of the Stretched SI-SNR loss is unclear; the phrase \"doubling the period of SISNR\" does not explain how Eq. (2) is derived from the standard SI-SNR, and the formula as written does not equal the conventional -10 log10(1 - cos^2(beta)). Please provide a derivation or a precise citation.","section":"Equation (2)"},{"comment":"The text says \"For L1 and L3 in Figure 2, we set them as 8000 and 3200 sampling points,\" but Figure 2 does not label which length corresponds to which panel; please state explicitly which truncation length is used for the convolution prompt in fusion (d).","section":"4.2, Figure 2"},{"comment":"The sentence \"the initial frame is discarded\" is confusing; please clarify which feature is concatenated with the far-end and microphone signals and what exactly is discarded after processing.","section":"3.2(b)"},{"comment":"The column layout is difficult to parse; please clarify whether, for each test-RIR type, the three metrics apply to both the DT and ST FE scenarios or whether PESQ/SDR are reported for DT and ERLE for ST FE, and label the columns accordingly.","section":"Table 1"},{"comment":"The denoised RIR spectrum appears visually different from the clean one; add a quantitative measure, such as spectral distance or RIR SNR, to support the claim that the masking module preserves RIR information.","section":"Figure 4"},{"comment":"Reference [33] points to a URL for the ICASSP 2023 AEC Challenge; please cite the specific dataset release and version so that the training and test mixtures are precisely identified.","section":"Reference [33]"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the oracle-RIR leakage in fusion (d) and the absence of any noisy-RIR experiment, despite the paper's own motivation. If the authors add ablations with corrupted or stale RIR prompts and report significance, the paper could become acceptable. I do not see a circularity problem in the derivation, but the evaluation protocol needs to match the deployment scenario described in the introduction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is real: instead of estimating the echo path from signals, measure the RIR and hand it to the network as a prompt. The cleanest variant (d) convolves the denoised, truncated RIR with the far-end signal and feeds the resulting synthesized echo as an extra input. That variant gives consistent gains on mismatched synthetic and real RIR test sets across two baselines, with the MTFAA SDR jump from 7.75 to 10.16 on real RIR being the standout number. The paper also explores three other fusion methods, reports compute cost, and promises code. That is solid, publishable work in the AEC subfield.\n\nThe soft spot is exactly what the stress-test note flags. Every reported gain uses a clean, oracle-level RIR prompt. The paper motivates the method with noisy pulse collection, shows a noisy-RIR spectrum, and argues traditional methods fail on noisy RIR — but it never runs an experiment where the RIR prompt itself is noisy or perturbed. The denoiser is described as a simple masking module, and its output quality is never quantified. If the denoiser is imperfect, or the measured RIR is stale because people moved, the prompt echo becomes biased and the Table 1 gains could shrink or reverse. That is a load-bearing gap, not a minor omission. The abstract overclaims when it says the method works in real environments where the RIR must be measured.\n\nThe other limitations are proportionate. Three of the four fusion variants underperform on the mismatch set, and variant (d) was selected after seeing results on all test sets, so there is a selection effect. No error bars or significance tests are reported. The \"information leakage\" point is real but not disqualifying: the method is explicitly designed to give the model the linear echo path, and that is legitimate when the RIR is accurate. It just means the gains partly measure the value of a true RIR, not the robustness of the denoising pipeline.\n\nIf I were refereeing this, I would ask for one experiment: add noise to the RIR prompt at a level resembling Figure 4b, run the full denoise-and-prompt pipeline, and report the metrics. Without that, the central claim is conditional. The clean-RIR result is still a useful contribution and worth publishing as a first step.\n\nRecommendation: send it to peer review, but with a clear request for the noisy-RIR experiment. It merits a serious referee, and the current version needs that experiment before the generalization claim can stand.","headline":"Using a measured RIR as a prompt is a genuine idea with clean-RIR gains, but the paper never tests the noisy-RIR scenario it motivates, so the headline result is conditional on an unvalidated denoising step.","tokens_in":8905,"tokens_out":2339,"would_cite":true,"duration_ms":45534,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Feeding a measured room impulse response to an acoustic echo canceller as a prompt improves generalization to unseen rooms.","keywords":["acoustic echo cancellation","room impulse response","RIR prompt","model generalization","deep learning","speech enhancement","prompt fusion"],"falsifier":"Measure AEC quality with method (d) while systematically degrading the measured RIR (add noise, truncate further, use an RIR from a different room, or simulate a moving speaker). If PESQ/SDR on the real-RIR test set do not improve over the no-prompt baseline once the RIR deviates by a modest amount, the central claim that measured RIRs are useful prompts is falsified. Conversely, ablating the denoiser and comparing results with ground-truth clean RIRs would quantify how much of the gain depends on denoising.","tokens_in":7891,"feed_emoji":"🎧","tokens_out":5780,"duration_ms":48701,"temperature":0.7,"pith_summary":"Data-driven acoustic echo cancellation models trained on synthetic data tend to degrade in real rooms, where the echo path is not directly observable. This paper claims that feeding the model a measured room impulse response (RIR) as an auxiliary prompt restores much of that lost performance. Among four fusion designs, the strongest is simple: denoise the RIR, truncate it, convolve it with the far-end signal to synthesize a 'prompt echo,' and pass that alongside the microphone and far-end inputs. On mismatched synthetic RIRs and real recorded RIRs, the prompt raises PESQ, SDR, and ERLE for both the ICCRN and MTFAA backbones, with a modest increase in computation. If this holds, devices that can re-measure their acoustic environment can keep neural echo cancellers accurate as rooms change.","feed_headline":"RIR prompt lifts echo cancellation in unseen rooms","feed_subtitle":"Denoised room impulse response convolved with the far-end signal boosts PESQ and SDR on real-room tests.","key_machinery":"The central object is the RIR prompt, a measured room impulse response converted into an auxiliary network input. Fusion method (d), the effective variant, consists of a simple masking-based denoiser, a truncation to 3200 samples, and a convolution of the denoised truncated RIR with the far-end signal to synthesize the prompt echo v'(n). This prompt echo is concatenated with the far-end and microphone signals in the time-frequency domain before the AEC backbone processes them. The mechanism is that v'(n) approximates the true acoustic echo v(n), giving the network a direct, informed target for subtraction rather than forcing it to infer the echo path from the far-end signal alone. The four fusion designs test different ways of packaging the RIR: learned embeddings, temporal averaging, cropped STFT frames, and the convolution-based prompt echo.","core_discovery":"The paper's central claim is that a room impulse response, used as a training-time and inference-time prompt, improves AEC generalization. Concretely, fusion method (d) preprocesses the measured RIR with a lightweight masking denoiser, truncates it to 3200 samples, and convolves it with the far-end signal x(n) to form a prompt echo v'(n). The AEC network then takes the far-end, microphone, and prompt echo as inputs to predict the near-end signal. In experiments on a 1000-sample real-RIR test set, this raises ICCRN PESQ from 2.05 to 2.19 and SDR from 3.61 to 4.64, and MTFAA SDR from 7.75 to 10.16, while adding less than one Giga-MAC to ICCRN's 1.93 G-MAC baseline. The authors argue the prompt echo lets the network exploit RIR information directly, and that even a noisy RIR, unusable for classical adaptive filters, still helps a learned model.","pith_inferences":["A natural extension is to test how robust the gains are to RIR estimation error: as the measured RIR deviates from the true one, the prompt echo diverges from the real echo, and there should be a noise level beyond which method (d) underperforms the no-prompt baseline.","The prompt-echo idea may transfer to other room-dependent speech tasks, such as dereverberation or target speaker extraction, where a measured RIR could be convolved with a reference signal to condition the network on the environment.","The method assumes a static echo path between pulse measurements; combining it with continuous RIR tracking or online adaptation could cover moving speakers and rotating devices."],"forward_implications":["If a device can measure and denoise its own RIR, a neural AEC can be adapted to a new room without retraining on that room's data.","Method (d) works across backbones: both ICCRN and MTFAA gain on mismatched and real RIR sets, suggesting the prompt is architecture-agnostic.","The gains are largest where generalization is hardest: on real recorded RIRs and on mismatched microphone-loudspeaker distances, the conditions synthetic-only models fail.","The added cost is small: ICCRN parameter count grows from 0.463M to 0.611M and MACs from 1.93G to 2.77G with method (d)."],"supporting_citations":[{"why":"Defines the ICCRN backbone whose generalization the RIR prompt improves.","marker":"[26]"},{"why":"Defines the MTFAA backbone; method (d) nearly doubles its real-RIR SDR.","marker":"[17]"},{"why":"Provides the image method used to simulate training and matched synthetic RIRs.","marker":"[34]"},{"why":"Supplies the real recorded RIRs used in the generalization test set.","marker":"[35]"},{"why":"Supplies the near-end and far-end speech dataset from the ICASSP 2023 AEC challenge.","marker":"[33]"},{"why":"Source of the S-SISNR time-domain loss used in training.","marker":"[28]"},{"why":"One of the two sources of the 'RI+Mag' complex-spectrum loss.","marker":"[30]"},{"why":"The other source of the 'RI+Mag' loss used to recover the complex spectrum.","marker":"[31]"}],"fun_headline_variants":["RIR prompt boosts AEC in unseen rooms","Denoised RIR prompt sharpens echo cancellation","Room impulse response as prompt improves AEC","RIR-guided AEC generalizes to real rooms","Prompt echo from RIR lifts AEC performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that at inference time the device can measure and denoise its room impulse response accurately enough that convolving it with the far-end signal yields a prompt echo close to the actual echo; if the measured RIR is noisy, outdated, or distorted by moving people or nonlinear speakers, the gains may disappear.","fun_headline_variants_meta":{"raw":{"variants":["RIR prompt boosts AEC in unseen rooms","Denoised RIR prompt sharpens echo cancellation","Room impulse response as prompt improves AEC","RIR-guided AEC generalizes to real rooms","Prompt echo from RIR lifts AEC performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1390,"prompt_tokens":873,"completion_tokens":517,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":445}},"tokens_in":489,"tokens_out":517,"duration_ms":4861,"temperature":1.0,"reasoning_tokens":445,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:13:24.600220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure AEC quality with method (d) while systematically degrading the measured RIR (add noise, truncate further, use an RIR from a different room, or simulate a moving speaker). If PESQ/SDR on the real-RIR test set do not improve over the no-prompt baseline once the RIR deviates by a modest amount, the central claim that measured RIRs are useful prompts is falsified. Conversely, ablating the denoiser and comparing results with ground-truth clean RIRs would quantify how much of the gain depends on denoising.","supporting_citations":[{"cited_title":"A conformer- based waveform-domain neural acoustic echo canceller optimized for asr accuracy","cited_arxiv_id":null,"evidence_quote":"Defines the ICCRN backbone whose generalization the RIR prompt improves."},{"cited_title":"Nonlinear acoustic echo cancellation based on volterra filters,","cited_arxiv_id":null,"evidence_quote":"Defines the MTFAA backbone; method (d) nearly doubles its real-RIR SDR."},{"cited_title":"Funnel deep complex u- net for phase-aware speech enhancement,","cited_arxiv_id":null,"evidence_quote":"Provides the image method used to simulate training and matched synthetic RIRs."},{"cited_title":"Conv-tasnet: Surpassing ideal time- frequency magnitude masking for speech separation,","cited_arxiv_id":null,"evidence_quote":"Supplies the real recorded RIRs used in the generalization test set."},{"cited_title":"Room impulse response generator,","cited_arxiv_id":null,"evidence_quote":"Supplies the near-end and far-end speech dataset from the ICASSP 2023 AEC challenge."},{"cited_title":"Conversational artificial intelligence in the aec industry: A review of present status, challenges and oppor- tunities,","cited_arxiv_id":null,"evidence_quote":"Source of the S-SISNR time-domain loss used in training."},{"cited_title":"Low- complexity acoustic echo cancellation with neural kalman filter- ing,","cited_arxiv_id":null,"evidence_quote":"One of the two sources of the 'RI+Mag' complex-spectrum loss."},{"cited_title":"NeuralKalman: A Learnable Kalman Filter for Acoustic Echo Cancellation","cited_arxiv_id":"2301.12363","evidence_quote":"The other source of the 'RI+Mag' loss used to recover the complex spectrum."}],"review_version":1}