{"id":"4344b5d3-8a08-4328-b7d6-a470a8f32f8d","arxiv_id":"2603.23723","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Feeding a spatial filter's cleaned output back into Kalman or particle filters dramatically improves moving-speaker tracking with almost no added computation.","lead":"This paper shows that feeding a spatially selective filter's cleaned output back into lightweight Kalman or particle filters substantially improves tracking of moving speakers, with almost no added computation. The result makes real-time target-speaker extraction in dynamic, multi-speaker rooms feasible with cheap Bayesian trackers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Autoregressive gain depends on treating enhanced SSF output as faithful clean observations in Eqs. (19)/(23); leakage or IPD distortion can make the feedback loop lock onto the wrong speaker, and only one SSF plus coarse Re-ACC is tested.","rationale":"I read the paper's central claim as: cheap Bayesian trackers, when fed the SSF's own enhanced output, can achieve near-neural tracking accuracy at negligible extra compute. The load-bearing assumption is that the enhanced output is a trustworthy observation of the target in the tracker's likelihood. This is exactly what Eq. (23) formalizes: Y_tk = d_tk S_tk + V_tk, with Ŝ_tk substituted for S_tk. If the SSF suppresses the target's inter-channel phase structure or leaks the interferer, the residual V̂_tk in Eq. (24) is not the intended noise, the recursive covariance in Eq. (25) becomes a biased function of θ̂, and the feedback is not a clean refinement loop but a potential lock-on to the wrong direction. The paper's own Fig. 4b shows that tuning α_IPD trades tracking accuracy against enhancement quality, and the MIMO-AR benefit is not uniform across trackers in Table I (e.g., PF MISO-AR is better than PF MIMO-AR), suggesting sensitivity to the spatial fidelity of the estimate. The lack of an oracle-substitution bound means the reported gains could partly reflect the SSF being trained to produce outputs that make the particular tracker's likelihood easy, rather than a general property of AR feedback. This does not invalidate the empirical results — the paper is careful, includes confidence intervals, and releases code/data — but it leaves the main mechanism underspecified and limits the claimed general compatibility with arbitrary SSFs. The real-world evidence is suggestive but too coarse to close this gap. I therefore agree with the reader's conditional assessment: the concern should be addressed before full acceptance, but the evidence does not warrant rejection.","tokens_in":19263,"tokens_out":8538,"duration_ms":101441,"concrete_test":"Run an oracle-substitution ablation on the synthetic test set: during MISO-AR and MIMO-AR inference, replace Ŝ_{t−1} in the tracker likelihoods (Eqs. 19 and 23) and in the covariance residual (Eq. 24) with the oracle clean target signal (reference-channel for MISO-AR, multichannel direct-path for MIMO-AR), leaving the SSF and all other pipeline components unchanged. Compare the resulting MAE/ACC with the reported AR rows in Table I. If the oracle-plugged MAE is within ~1° of the reported values, the plug-in approximation is not a first-order failure mode; if it is substantially better, processing degradation from the SSF is a major unquantified factor in the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that feeding the SSF's previous enhanced output back into Bayesian tracker likelihoods robustly improves DoA estimation and, in turn, enhancement — rests on the plug-in approximation in Sec. IV: the enhanced signals Ŝ_{1:t−1} are substituted for clean target signals S_{1:t−1} in the generative observation models (Eqs. 17–27). This is only valid if (a) the SSF suppresses the interfering speaker without distorting the target, and (b) for MIMO-AR, the multichannel estimate preserves the inter-microphone phase structure encoded in the steering vector d_k(θ_t). Neither condition is guaranteed. The SSF is trained on reconstruction losses (Eqs. 39–41), not on likelihood fidelity; the IPD loss is a soft auxiliary term with a tuned weight α_IPD, and Fig. 4b shows a clear MAE/PESQ tradeoff. If Ŝ leaks the competing speaker, the residual V̂ in Eq. (24) is biased, the recursive noise-covariance update Eq. (25) is mis-specified, and the tracker can lock onto the wrong source — an error the feedback loop can amplify rather than recover from. The paper tests only SpatialNet; no oracle-substitution or likelihood-fidelity experiment is reported. The real-world evaluation (Fig. 8) uses a coarse side-of-array Re-ACC, which cannot distinguish correct-speaker tracking from tracking the wrong speaker that is temporarily on the expected side. Therefore the empirical support does not yet rule out failure for less phase-preserving SSFs or more adverse acoustic conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes autoregressive (AR) guidance of deep spatially selective filters (SSFs) for moving-speaker extraction, by feeding the SSF's previously enhanced speech back into lightweight Bayesian target-speaker trackers (wrapped Kalman filter and bootstrap particle filter). Two AR schemes are presented: MISO-AR, where the enhanced single-channel estimate is incorporated into the tracker's likelihood, and MIMO-AR, where a multichannel SSF estimate replaces the noisy observation in the generative model. The paper also introduces a social-force-model-based synthetic dataset for realistic speaker trajectories. Experiments on synthetic and real-room recordings show that the AR methods substantially improve tracking accuracy and enhancement quality over the concatenative baseline, with negligible added compute, and achieve accuracy competitive with neural trackers at a fraction of the computational cost.","tokens_in":19832,"tokens_out":4907,"duration_ms":54149,"significance":"If the central claim holds, the paper makes a useful contribution: it shows that a cheap statistical tracker can exploit the output of a deep SSF to steer it accurately, potentially replacing much more expensive neural trackers in online causal pipelines. The paper also ships a dataset-generation framework, reports confidence intervals on synthetic metrics, includes real-room evaluations, and provides compute comparisons — all strengths. The main open risk is that the AR gain depends on treating the SSF's possibly degraded output as a faithful clean-speech observation inside the tracker, and the neural-tracker comparison is not controlled for the additional feedback information available only to the Bayesian AR methods.","major_comments":[{"comment":"The central AR mechanism relies on the plug-in substitution of enhanced speech \\hat{S} for clean speech S in the tracker likelihoods, while explicitly neglecting SSF processing degradation. This is load-bearing: if \\hat{S} leaks the interferer, the residual in Eq. (24) is biased and the recursive noise-covariance update in Eq. (25) is mis-specified; if the MIMO output corrupts phase structure, the steering-vector likelihood in Eq. (23) can lock onto the wrong direction. The paper acknowledges this limitation but does not bound it. Only SpatialNet is tested, and Fig. 4b shows a clear tradeoff between IPD preservation and reconstruction quality. I recommend adding an oracle-substitution experiment (feeding clean target speech or leakage-controlled mixtures into the same trackers) and evaluating at least one additional SSF or a deliberately phase-distorted variant, to demonstrate robustness","section":"Sec. IV-A, Eqs. (17)-(27) and Table I"},{"comment":"The comparison with neural trackers is not controlled for the AR feedback. The neural baselines (SELDnet, CNN/LSTM) receive only the noisy multichannel signal and the initial DoA, while the Bayesian AR methods additionally observe the SSF's enhanced output. Thus the comparison conflates tracker architecture with an extra source of information, and the claim that the Bayesian method matches or beats SELDnet at less than one-tenth the compute is only about non-AR neural trackers. A neural tracker embedded in the same AR loop might close the gap. Please either integrate the neural baselines into the AR loop (e.g., by feeding enhanced features) or explicitly restrict the claim to non-AR neural trackers, and avoid presenting the result as a pure architecture advantage.","section":"Sec. VII-B, Fig. 5"},{"comment":"The real-world DoA evaluation uses the regional accuracy (Re-ACC), a coarse side-of-array metric. This cannot distinguish correct target tracking from tracking an interferer who is momentarily on the expected side — precisely the failure mode associated with the leaky-feedback concern. Since the synthetic evaluation carries the main DoA claim, the real-world results should be framed primarily as supporting enhancement/generalization evidence rather than as exact DoA validation. I suggest adding a small number of manually annotated ground-truth positions or a finer event-based analysis around speaker crossings.","section":"Sec. VII-D, Fig. 8"},{"comment":"The fine-tuning procedure fixes tracker hyperparameters after the first epoch and performs a final parameter sweep for the AR configuration, but the paper does not state whether the concatenative baselines received an equally thorough hyperparameter optimization. If the Concat trackers were left at less-tuned settings, part of the observed gain in Table I could be due to tuning rather than the AR mechanism. Please clarify the hyperparameter search for all compared pipelines, including the neural trackers in Fig. 5.","section":"Sec. VI-B and Fig. 4a"}],"minor_comments":[{"comment":"Typos: \"Boostrap\" should be \"Bootstrap\" (Sec. VI-A), \"competetive\" should be \"competitive\" (Sec. VII-B), and \"real-word trajectories\" in the conclusion should be \"real-world trajectories\".","section":"Throughout"},{"comment":"The arrow labels on the y-axes (\"MAE [°]←\" and \"ACC [%]→\") are visually confusing; please use standard axis labels with a legend for the arrow direction.","section":"Fig. 5"},{"comment":"The Re-ACC definition could be stated more precisely in the text; currently it is described mainly via Fig. 7. A one-sentence mathematical definition would improve reproducibility.","section":"Sec. VII-D"},{"comment":"The social-force model parameters are reported in the text, but a single summary table listing all model constants and their sampling distributions would aid implementation.","section":"Sec. V-B"}],"recommendation":"major_revision","confidential_remarks":"This is a solid extension of the authors' earlier conference work, with useful synthetic data and a clear computational-cost story. The main risk is that the headline comparisons (AR vs. concatenative, Bayesian vs. neural) are not fully controlled, and the core plug-in approximation is not stress-tested. I recommend a major revision that adds controlled experiments and explicitly narrows the claims where needed. The paper is not a reject; the issues are fixable within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThe thing to know about this paper: it's a well-executed follow-up to the authors' own WASPAA/ICASSP work on self-steering spatial filters. They take two lightweight Bayesian trackers (wrapped Kalman, bootstrap particle filter) and feed the previous frame of the SSF's enhanced speech back into the likelihoods. That simple autoregressive loop cuts DoA error dramatically on synthetic data (e.g., Bootstrap PF MISO-AR MAE from 21.7° to 6.5°) and improves enhancement scores, at negligible added compute. They also release a social-force-model synthetic trajectory dataset, which is a genuinely useful resource for the moving-speaker tracking/enhancement community.\n\nWhat's actually new: the enhanced-magnitude IPD reweighting in Eq. 19 for the wrapped KF, and the bootstrap PF with the AR likelihood plus the EMA noise-covariance update (Alg. 1). The MIMO extension is from their prior work, so the new contribution is these Bayesian reformulations plus the dataset. The experimental work is solid: confidence intervals, cross-dataset training evaluation (Fig. 6), and real-room recordings with NISQA/WER. The paper ships code and the dataset, which is real evidence in favor of reproducibility.\n\nThe soft spots are in proportion. The main one: the comparison to neural trackers in Fig. 5 is not controlled. The neural baselines (SELDnet, CNN/LSTM) don't get the autoregressive feedback of the enhanced speech, so claiming \"competitive with neural trackers at a fraction of the cost\" is overstated — you're comparing AR Bayesian vs non-AR neural. An apples-to-apples comparison would require giving the neural trackers the same feedback or removing the AR loop from the Bayesian methods. That's a real flaw in the baseline design, but it doesn't undermine the core result that AR helps the Bayesian filters a lot.\n\nSecond, the plug-in of enhanced speech Ŝ as clean speech S in the likelihoods (Eqs. 17-27) is acknowledged as an approximation but not stress-tested. If the SSF leaks the interferer or distorts inter-microphone phase, the feedback can lock onto the wrong source. The paper only tests one SSF (SpatialNet), so we don't know whether the gain holds for other SSFs. An oracle-substitution experiment — replace Ŝ with true S, or with a deliberately degraded S — would have made the mechanism much clearer. In the real-room evaluation, the Re-ACC metric is coarse (side-of-array), so it can't distinguish tracking the correct speaker on the right side from tracking the wrong speaker on that side. The synthetic results with ground truth DoA partially mitigate this, but the real-world claim would be stronger with a finer metric.\n\nMinor: a handful of validation-tuned hyperparameters (particle count N, EMA coefficient, IPD loss weight) are not all reported, which slightly hurts reproducibility, though the released code likely covers them.\n\nOverall, I think the central claim holds up: autoregressive guidance of lightweight Bayesian trackers is a cheap way to improve moving-speaker extraction. The paper is not a breakthrough, but it's a useful, honest engineering contribution with reproducible components. It deserves a serious referee. The authors should be asked to fix the neural baseline comparison and add an oracle-substitution test for the plug-in assumption, but neither should be a desk-reject reason.\n\nI'd bring it to reading group and cite it for the dataset and the AR formulations.","headline":"Solid extension of the authors' self-steering SSF line: autoregressive feedback into lightweight Bayesian trackers genuinely helps, but the neural baseline comparison is uncontrolled and the plug-in of enhanced speech as clean observations is the main risk.","tokens_in":20231,"tokens_out":3049,"would_cite":true,"duration_ms":30713,"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":"The paper claims that a moving speaker's direction can be tracked by feeding the spatial filter's own previous output back into a Kalman or particle filter, achieving accuracy competitive with neural trackers at negligible added compute.","keywords":["moving speaker tracking","direction of arrival estimation","Bayesian filtering","Kalman filter","particle filter","autoregressive guidance","spatially selective filter","target speaker extraction"],"falsifier":"Take a dataset with exact ground-truth DoA and a second independently pretrained spatial filter of different architecture. If, during a sustained speaker crossing, the AR tracker's error exceeds the concatenative tracker's, or if deliberately corrupting the fed-back signal (e.g., replacing it with the interferer's enhanced output) does not degrade tracking, then the plug-in clean-speech assumption is refuted.","tokens_in":19194,"feed_emoji":"🎙️","tokens_out":10266,"duration_ms":89490,"temperature":0.7,"pith_summary":"This paper argues that a moving speaker's direction can be tracked accurately by closing the loop between a deep spatial filter and a cheap Bayesian tracker: after the filter produces its first enhanced estimate, that cleaned signal is fed back into a Kalman or particle filter as an observation, instead of the tracker relying only on the raw microphone mixture. With nothing but the speaker's initial direction as guidance, this autoregressive feedback reduces direction-of-arrival error substantially and improves the quality and intelligibility of the extracted speech. The method is compatible with arbitrary spatial filters, adds negligible computation, and generalizes to real rooms with reverberation up to 800 ms. A reader should care because it offers a way to replace expensive neural tracking networks with lightweight recursive filters in real-time hearing and communication devices.","feed_headline":"Reusing the filter's output cuts moving-speaker DoA error threefold","feed_subtitle":"Feeding cleaned speech back into Kalman and particle filters shrinks mean direction error from 21.7° to 6.5°.","key_machinery":"The load-bearing mechanism is temporal feedback in a frame-wise causal pipeline: at time t, the previously enhanced speech estimate replaces or conditions the tracker's likelihood in place of the raw microphone observation. A MIMO variant of the spatial filter is introduced whose output retains inter-microphone phase structure, so the enhanced signal can act as a multichannel pseudo-observation; a MISO variant instead uses the single-channel estimate as a frequency selector that emphasizes target-dominated bins in the DoA estimator. The recursive Bayesian update then couples the filter and tracker into a closed loop whose added cost is only a few million multiply-accumulate operations per se","core_discovery":"The central claim is that the enhanced output of a spatially selective filter can serve as the observation model inside a recursive Bayesian tracker, creating a self-steering loop. The paper develops two ways to do this: MISO-AR, which uses the single-channel cleaned speech to reweight frequency bins in a wrapped-Gaussian Kalman likelihood or to condition a particle-filter likelihood, and MIMO-AR, which modifies the spatial filter to output a multichannel estimate that can replace the raw mixture in the likelihood. In simulation, the bootstrap particle filter in the MISO-AR configuration reduces mean absolute DoA error from 21.65° to 6.47° and raises 10°-accuracy from 56.2% to 87.6%, while i","pith_inferences":["The same plug-in principle could be extended to other target cues—elevation, range, or binaural hearing-aid configurations—wherever a causal filter output is available; this is an inference, not tested here.","The success of the single-channel MISO-AR particle filter, without explicit spatial-cue preservation, suggests the main benefit of feedback is a cleaner, more target-dominated input to the likelihood rather than restored inter-microphone phase; a testable extension would substitute a single-channel denoiser's output to isolate that effect.","The real-room evaluation uses a coarse side-of-array metric rather than exact DoA ground truth; an independent motion-captured dataset could determine whether the AR gains reflect true direction accuracy or mainly reduced speaker confusion after crossings.","The new trajectory simulator may be useful as a training distribution for other moving-source tasks, and its public release invites cross-dataset comparisons beyond the tested room sizes and array geometry."],"forward_implications":["In the strongest simulated configuration (bootstrap particle filter with MISO-AR), mean DoA error falls from 21.65° to 6.47° and 10°-accuracy rises from 56.2% to 87.6%, with PESQ improving from 1.93 to 2.04.","The autoregressive Kalman filter, which by itself cannot handle moving speakers (MAE above 30°), drops to 6.65° once the spatial filter is switched to its multichannel MIMO-AR form.","The approach costs almost nothing beyond the base filter: trackers require 0.3–2.5 MMAC/s, and the MIMO extension adds fewer than 500 parameters and about 800 kMAC/s per kHz of bandwidth.","AR-guided Bayesian trackers reach accuracy comparable to a reference neural tracker while using an order of magnitude less computation.","Spatial filters trained on pedestrian-like moving-source trajectories retain quality on unseen motion patterns, including real-world walking data."],"fun_headline_variants":["Self-steering audio filter tracks moving speakers with less error","Feeding cleaned speech back into tracker cuts DoA error by 70%","Autoregressive feedback improves moving-speaker localization in real time","Filter output as tracker input: a self-correcting loop for moving speakers","Bayesian tracker slashes DoA error for moving speakers using filter feedback"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The results depend on treating the spatial filter's previous enhanced output as a faithful clean-speech observation inside the tracker's likelihoods, even though the filter was trained on oracle directions and its errors are never modeled; if that output leaks the interfering speaker, the feedback loop can lock onto the wrong direction.","fun_headline_variants_meta":{"raw":{"variants":["Self-steering audio filter tracks moving speakers with less error","Feeding cleaned speech back into tracker cuts DoA error by 70%","Autoregressive feedback improves moving-speaker localization in real time","Filter output as tracker input: a self-correcting loop for moving speakers","Bayesian tracker slashes DoA error for moving speakers using filter feedback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":2961,"prompt_tokens":715,"completion_tokens":2246,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":2163}},"tokens_in":459,"tokens_out":2246,"duration_ms":15569,"temperature":1.0,"reasoning_tokens":2163,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:32:27.220930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with exact ground-truth DoA and a second independently pretrained spatial filter of different architecture. If, during a sustained speaker crossing, the AR tracker's error exceeds the concatenative tracker's, or if deliberately corrupting the fed-back signal (e.g., replacing it with the interferer's enhanced output) does not degrade tracking, then the plug-in clean-speech assumption is refuted.","supporting_citations":[],"review_version":1}