{"id":"d3656d6f-fb05-4db8-be9b-16dbfbdeba0f","arxiv_id":"2507.16104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A windowed cross-attention module outperforms the transform-average-concatenate module for speech enhancement with asynchronous, arbitrarily placed microphones.","lead":"This paper introduces a windowed cross-attention module that lets neural speech enhancement models handle microphones with different time delays and clock drifts. It reports better predicted speech-quality scores than the standard TAC module on two enhancement architectures in simulated ad-hoc meeting scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Clock-drift simulation is too weak to test the claimed robustness: over 10 s, N(16000,0.5) Hz drift accumulates <1 ms, two orders below the 90 ms attention window.","rationale":"The reader's weakest-assumption analysis and mine converge: the simulated clock drift is orders of magnitude too small to test the mechanism the paper claims to handle. I agree with the conditional verdict because the WCA-vs-TAC comparison for bounded fixed latency is internally consistent—with a 90 ms window and ≤80 ms offsets, WCA should win—and the reported gains in Table 2 are plausible. The unresolved issue is external validity: the paper's motivation (Section 1) cites >100 ms network latency and clock drift as real-world problems, but the evaluation never pushes either beyond the fixed window. No formal verification or code is provided; the demo page is illustrative rather than a controlled test, and the listed baseline SAMbA is not benchmarked. The strongest single check is to lengthen the test clips and increase drift so accumulated misalignment approaches the window size; this would directly test whether the attention window tracks drift or merely tolerates small offsets. I would not reject the paper: the architectural idea is clear, and the conditional framing matches the evidence.","tokens_in":9785,"tokens_out":6444,"duration_ms":69275,"concrete_test":"Re-run the CRUSE+WCA vs CRUSE+TAC evaluation on 10-minute clips with per-device sample-rate offsets of 100 ppm (realistic clock drift, accumulating to ~60 ms per stream and up to ~120 ms between streams by the end), keeping the 90 ms window unchanged, and compare OVRL/XLSR in the first vs last 30 s. If WCA's advantage over TAC collapses after accumulated drift exceeds the window, the clock-drift robustness claim is refuted for realistic meeting durations; if it persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that WCA is robust to clock drift and generalizes to real asynchronous meetings rests on the simulation in §5.1: per-device sample rates drawn from N(16000, 0.5) Hz and fixed delays in [−40, 40] ms, with 10 s test clips. A 0.5 Hz offset at 16 kHz is 31 ppm; over 10 s this accumulates at most ~0.3 ms per stream, and even the 2 Hz ablation reaches only ~1.25 ms. The 90 ms attention window (L=4, §5.2) is therefore never stressed by drift. The latency range is also fully covered by the window (max pairwise offset 80 ms < 90 ms), so the experiments do not probe the window boundary or the >100 ms latency cited as motivation in §1. Consequently, Table 2 only shows that WCA handles small fixed offsets, not that it tracks drift over time or generalizes to longer meetings; the §5.4 statement that the model 'demonstrates robustness to clock drift' is not supported by the reported conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a windowed cross-attention (WCA) module for multi-microphone speech enhancement in distributed asynchronous settings. WCA replaces the time-index-aligned TAC aggregation with a fixed temporal cross-attention window that can align features across devices with unknown latency and clock drift. The module is integrated into two backbones, iFaSNet and CRUSE, and evaluated on simulated noisy reverberant mixtures with multiple speakers, per-microphone time shifts, and per-device sample-rate offsets. Three training targets are compared: a fixed random microphone, the minimum-latency microphone, and the microphone closest to each active speaker. The reported results (DNSMOS, XLSR-MOS, and cepstral distance) show WCA generally outperforming TAC, with the closest-microphone target giving the largest gains. The paper also presents an ablation varying drift strength, delay magnitude, and speaker overlap, and concludes that WCA is robust to clock drift and generalizes to real asynchronous meeting conditions.","tokens_in":9938,"tokens_out":4080,"duration_ms":45528,"significance":"If the claims were fully supported, the WCA module would be a useful contribution: it is permutation- and microphone-count-invariant, has bounded memory cost through the fixed window, and can be dropped into existing enhancement backbones. The comparison with TAC under matched architectures is a fair and informative control, and the use of two independent backbones plus external learned MOS predictors strengthens the main architectural comparison. The closest-microphone training target for multi-talker mixtures is also a sensible idea that yields consistent improvements. However, the headline robustness-to-clock-drift claim is not actually exercised by the simulated conditions, because the accumulated drift over 10 s clips is sub-millisecond and the entire simulated latency range lies inside the attention window. The paper also lacks error bars or significance tests, and the best target strategy is selected on the same test set used for the final headline numbers. These issues leave the general claims about real asynchronous meetings unsupported, though the core WCA-versus-TAC comparison for small fixed offsets remains plausible.","major_comments":[{"comment":"The clock-drift simulation is too weak to test the claimed robustness. Per-device sample rates are drawn from N(16000, 0.5) Hz, which is a 31 ppm offset; over a 10 s clip this accumulates at most roughly 0.3 ms of relative drift, and even the ablation with std 2 Hz accumulates only about 1.25 ms. These values are two orders of magnitude smaller than the configured 90 ms attention window, so Table 2 shows only that WCA handles small fixed offsets, not that it tracks time-varying drift. The statement in §5.4 that the model 'demonstrates robustness to clock drift' is not supported by the reported conditions; experiments with larger sample-rate offsets, longer utterances, or time-varying drift are needed.","section":"§5.1, Table 2"},{"comment":"The simulated latency range is fully covered by the attention window, so the window boundary is never stressed. Delays are sampled between -40 ms and 40 ms, giving a maximum pairwise offset of 80 ms, while L=4 is described as a 90 ms cross-attention window. The >100 ms network latency cited as motivation in §1 is therefore never tested. The conclusion that WCA handles large device latency requires either an experiment with offsets exceeding the window size or a principled argument that the window can be enlarged without performance loss.","section":"§5.1, §5.2"},{"comment":"No error bars, confidence intervals, or repeated-seed results are reported for any of the DNSMOS, XLSR-MOS, or CD numbers. Many of the reported advantages are small (for example, OVRL differences of 0.1-0.2), and it is not clear whether these are stable across training runs or test-set subsamples. The claim of 'consistent' improvement should be backed by variance estimates or a significance test, or at least by an explicit statement that the tables report a single run.","section":"Tables 1 and 2"},{"comment":"Two claims about Table 1 need correction or stronger support. First, iFaSNet with the random-microphone target does not consistently favor WCA: OVRL is 2.27 for WCA versus 2.28 for TAC, and SIG is 2.67 versus 2.68, so the text that says 'only a slight improvement over TAC' is not accurate for those metrics. Second, the closest-microphone target is declared the best among the three strategies after observing test-set results, and the headline OVRL/XLSR numbers are then reported for that same test set; this target-selection step should be performed on a validation set or otherwise accounted for to avoid optimistic selection bias.","section":"§5.4, Table 1"}],"minor_comments":[{"comment":"The abstract calls the closest-microphone target an 'optimal training target,' but the paper only shows it is empirically best among three strategies on one test set; this wording overstates the finding unless a theoretical argument or broader evaluation is provided.","section":"Abstract and §4.3"},{"comment":"There are several typos and grammatical slips, including 'awkard' in §1, 'microhpone' in §2, and 'synchonization' in §3; these should be corrected.","section":"§1, §2, §3"},{"comment":"The softmax normalization in the windowed attention is not fully specified: it is unclear whether the softmax is taken over the 2L+1 window within each microphone pair n, or jointly across all n and the window. This matters for the definition of the attention weights and should be stated unambiguously.","section":"Eq. (9)"},{"comment":"The mapping from L=4 to a '90 ms' window is not derived. With a 20 ms STFT window and 50% overlap, the hop size is 10 ms, and a window spanning i-L to i+L covers 80 ms of offsets plus frame boundary effects; the paper should state the exact relationship between L, hop length, and the physical offset range in milliseconds.","section":"§5.2"},{"comment":"The claim that the method 'generalized well to real-world conditions' is supported only by an unquantified website of audio examples. If a real-recording evaluation exists, it should be reported with metrics; otherwise the claim should be tempered to 'anecdotally demonstrated on audio examples.'","section":"§5.4, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is likely appropriate for a workshop-style publication, and the WCA architecture is a reasonable incremental contribution. The central problem is that the clock-drift claim is untested under the simulated conditions, and the latency regime does not reach the window boundary. The absence of variance information and the test-set-based choice of the best target strategy further weaken the headline numbers. These issues are fixable within the scope of the manuscript, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Get to read this. Short version: WCA is a reasonable idea and there is a real, if modest, result here: on simulated fixed offsets up to 40 ms, swapping TAC for WCA helps both iFaSNet and CRUSE, especially with the closest-microphone target. I'd trust that comparison more than the abstract does. The module is permutation/cardinality-invariant, cheap to add, and the windowed attention over time is a genuine extension of TAC and graph attention, which only aggregate same-time-index frames.\n\nThe soft spot is exactly the one the stress-test flagged. The drift simulation uses per-device sample rates drawn from N(16000, 0.5) Hz, which over 10 s clips gives sub-millisecond relative drift; even the std=2 ablation is about 1.25 ms. The 90 ms window is never stressed. The fixed-delay condition max pairwise 80 ms is also inside the window. So Table 2 shows WCA can cope with small fixed offsets, not that it tracks drift over time. The sentence in Section 5.4 calling the model robust to clock drift is not supported. If they reframe the contribution as 'robust to bounded, slowly-varying offsets within a fixed window,' that's honest and still useful.\n\nOther soft spots are smaller. No error bars or significance tests, and the metrics are learned MOS predictors; given the effect sizes (0.1–0.5 OVRL), some variance reporting is needed. The CM target is oracle (known geometry) and they pick the best target on the test set, so the 'optimal target' claim is inflated. SAMbA is mentioned but not benchmarked, and the faster-convergence claim has no curve. The iFaSNet random-target row shows only a tiny gain, so 'consistently outperforms' is mostly true but not uniformly.\n\nWho's this for? People working on ad-hoc microphone arrays and neural beamforming. It is a legitimate workshop-or-conference contribution as a module, not a paradigm shift. I'd send it to review—the core architecture and comparison deserve referee time—but insist the authors either run a real drift protocol or soften the claim, add error bars, and add SAMbA or an explicit-alignment baseline.\n\nRecommendation: engage, with revisions. Not a desk reject.","headline":"A sensible attention module and a fair latency comparison, but the clock-drift claim is not exercised by the simulation—the contribution is narrower than the title suggests.","tokens_in":10515,"tokens_out":2396,"would_cite":true,"duration_ms":27567,"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":"A windowed cross-attention module lets speech enhancement networks align and aggregate asynchronous microphone streams, outperforming the same-time-index TAC module on both tested backbones.","keywords":["asynchronous microphone arrays","clock drift","cross-attention","speech enhancement","multi-talker","ad-hoc microphone array","device latency","transform-average-concatenate"],"falsifier":"Record a multi-minute real meeting with consumer devices whose clocks drift far enough that the relative offset between two streams exceeds 90 ms, or with bursts of packet loss, and compare WCA against TAC; if the WCA advantage disappears or reverses once the misalignment exceeds the fixed window, the claimed drift robustness is not general.","tokens_in":9501,"feed_emoji":"🎤","tokens_out":14219,"duration_ms":134575,"temperature":0.7,"pith_summary":"Most multi-microphone speech enhancement assumes the microphone streams are time-synchronized; in a real meeting each device records on its own clock, so the streams arrive with unknown latency and gradually drift apart. This paper proposes a windowed cross-attention (WCA) module that aligns features across microphones by letting each time frame attend to a small neighborhood of frames in every other microphone, instead of averaging the same time index the way the standard transform-average-concatenate (TAC) module does. Inserting WCA into two existing enhancement networks and training toward the closest microphone to each active speaker yields better perceptual quality than TAC under simulated latency and clock drift, with the best configuration reaching an overall DNSMOS quality score (OVRL) of 2.41 versus 1.92 for TAC. The practical stake is that meeting capture could use every device in the room without manual muting, explicit synchronization, or a fixed array geometry.","feed_headline":"Cross-attention window beats same-time pooling for drifting mics","feed_subtitle":"WCA aligns unknown device latency and clock drift, and the closest-microphone target gives the largest quality gains.","key_machinery":"The central object is the Windowed Cross-Attention (WCA) layer, a cross-attention computation restricted in time. For each frame $i$ of microphone $m$, the query $Q_m[i]$ attends only to keys and values drawn from frames $j$ of every other microphone $n$ with $|i-j| \\le L$; with $L=4$ encoded frames this is a 90 ms window in the experiments. The output is $A_m[i] = \\sum_n \\mathrm{softmax}\\big(Q_m[i] \\cdot K_n^u[i]^\\top / \\sqrt{d}\\big) \\, V_n^u[i]$, where $K_n^u[i]$ and $V_n^u[i]$ are the unfolded local windows of microphone $n$. Restricting attention in this way reduces the memory cost from $O(M^2 T^2)$ to $O(M^2 T L)$, allows gradual drift to be tracked within the window, and makes the module invariant to microphone permutation and to the number of microphones. The second load-bearing piece is the training target: the target is not a fixed reference microphone but the sum of direct-path signals from the microphone closest to each active speaker, which the paper argues is the better objective for multi-talker scenes.","core_discovery":"The paper's central claim is that a temporal cross-attention restricted to a fixed local window can do the synchronization that explicit alignment methods struggle with. Under per-device latency up to 40 ms and clock drift modeled as per-device sample-rate offsets, models equipped with windowed cross-attention (WCA) beat transform-average-concatenate (TAC) models across nearly all configurations, and the largest gains come from training toward the closest microphone to each active speaker. On the CRUSE backbone with that target, WCA reaches a DNSMOS overall score (OVRL) of 2.41 and an XLSR-MOS score of 2.35, against 1.92 and 2.15 for TAC; WCA also outperforms the single-channel baseline with PickNet selection. The paper interprets this as evidence that WCA is a reusable, permutation- and number-invariant module for asynchronous multi-microphone enhancement, and that the closest-microphone-per-speaker target is the better training objective for multi-talker scenes.","pith_inferences":["The fixed 90 ms window sets an upper bound on the inter-device offset the model can absorb; for long recordings or heavy jitter, a wider window, an external drift-compensation front-end, or adaptive window sizing would be needed, none of which the paper tests.","The closest-microphone-per-speaker target is oracle-informed: forming it requires knowing which microphone is nearest to each active speaker at every moment, so a practical deployment would need a separate localization or speaker-diarization front-end to realize the same gains.","The reported metrics are non-intrusive perceptual scores that tolerate small time offsets, so the results demonstrate perceived-quality improvement but not necessarily sample-accurate synchronization; a downstream automatic speech recognition evaluation would be a stricter test of alignment quality.","The same windowed cross-attention mechanism could in principle align other unsynchronized sensor streams with bounded skew, such as distributed hearing-aid or wearable microphones, because the module only assumes a bounded temporal offset and permutation invariance."],"forward_implications":["Deploying WCA in an enhancement model means meeting capture can use all nearby devices simultaneously instead of muting all but one, because the model learns to align the asynchronous streams itself.","A model trained once with WCA is not tied to a fixed number or arrangement of microphones, since the attention is invariant to microphone permutation and count.","The closest-microphone-per-speaker training target improves perceived quality under multi-talker overlap, so systems that know which device is closest to each speaker can train toward a cleaner target than a fixed reference channel.","Faster convergence with WCA suggests the module also eases optimization relative to TAC, not just final quality."],"supporting_citations":[{"why":"Supplies TAC, the same-time-index pooling module that WCA replaces and outperforms under asynchronous input.","marker":"[8]"},{"why":"Provides iFaSNet, one of the two enhancement backbones into which WCA is integrated and tested.","marker":"[9]"},{"why":"Provides CRUSE, the second backbone whose bottleneck is augmented with WCA and which yields the best reported scores.","marker":"[39]"},{"why":"Gives the PickNet channel-selection baseline that fails to handle multi-talker asynchronous input and is outperformed by WCA-equipped CRUSE.","marker":"[25]"},{"why":"Defines DNSMOS P.835, the non-intrusive perceptual metric producing the SIG, BAK, and OVRL results.","marker":"[42]"},{"why":"Defines XLSR-MOS, the self-supervised quality metric used alongside DNSMOS in the main results.","marker":"[43]"},{"why":"Provides image-source room impulse response simulation used to create the reverberant noisy mixtures and the direct-path training target.","marker":"[37]"}],"fun_headline_variants":["Windowed cross-attention beats TAC for drifting mics","Async mic arrays? Windowed attention handles latency and drift","Closest-mic training target improves multi-talker enhancement","Permutation-invariant windowed attention for async mics","Windowed attention syncs unknown mic delays without alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim stands on the assumption that real-world clock drift and latency stay small enough to fall inside the fixed 90 ms attention window over the audio duration, with no large time-varying jitter or packet loss beyond what was simulated.","fun_headline_variants_meta":{"raw":{"variants":["Windowed cross-attention beats TAC for drifting mics","Async mic arrays? Windowed attention handles latency and drift","Closest-mic training target improves multi-talker enhancement","Permutation-invariant windowed attention for async mics","Windowed attention syncs unknown mic delays without alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001133,"raw_usage":{"total_tokens":4708,"prompt_tokens":948,"completion_tokens":3760,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":3679}},"tokens_in":564,"tokens_out":3760,"duration_ms":30513,"temperature":1.0,"reasoning_tokens":3679,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:18:10.487362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a multi-minute real meeting with consumer devices whose clocks drift far enough that the relative offset between two streams exceeds 90 ms, or with bursts of packet loss, and compare WCA against TAC; if the WCA advantage disappears or reverses once the misalignment exceeds the fixed window, the claimed drift robustness is not general.","supporting_citations":[{"cited_title":"End-to-end microphone permutation and number invariant multi-channel speech separation,","cited_arxiv_id":null,"evidence_quote":"Supplies TAC, the same-time-index pooling module that WCA replaces and outperforms under asynchronous input."},{"cited_title":"Implicit filter-and-sum network for end-to-end multi-channel speech separation,","cited_arxiv_id":null,"evidence_quote":"Provides iFaSNet, one of the two enhancement backbones into which WCA is integrated and tested."},{"cited_title":"Effect of noise suppression losses on speech distortion and ASR performance,","cited_arxiv_id":null,"evidence_quote":"Provides CRUSE, the second backbone whose bottleneck is augmented with WCA and which yields the best reported scores."},{"cited_title":"Picknet: Real-time channel selection for ad hoc microphone arrays,","cited_arxiv_id":null,"evidence_quote":"Gives the PickNet channel-selection baseline that fails to handle multi-talker asynchronous input and is outperformed by WCA-equipped CRUSE."},{"cited_title":"Dnsmos p. 835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,","cited_arxiv_id":null,"evidence_quote":"Defines DNSMOS P.835, the non-intrusive perceptual metric producing the SIG, BAK, and OVRL results."},{"cited_title":"Distillation and pruning for scalable self- supervised representation-based speech quality assessment,","cited_arxiv_id":null,"evidence_quote":"Defines XLSR-MOS, the self-supervised quality metric used alongside DNSMOS in the main results."},{"cited_title":"Image method for efficiently simulating small-room acoustics,","cited_arxiv_id":null,"evidence_quote":"Provides image-source room impulse response simulation used to create the reverberant noisy mixtures and the direct-path training target."}],"review_version":1}