{"id":"6e9e2816-ff21-41f4-aa39-1191ca49db2c","arxiv_id":"2504.14817","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A GRU-based neural network updated by the instantaneous-error gradient identifies time-varying HRIRs from a continuously rotating speaker array, achieving stable NM below -23 dB in simulation and experiment.","lead":"Researchers built a deep neural network that reconstructs head-related impulse responses while a speaker array rotates continuously, cutting measurement time. In simulations and a KEMAR experiment, it beat standard adaptive-filter methods by more than 7 dB in normalized misalignment.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sharper weakness is not one-sample underdetermination but that every reported NM/LSD comes from the same single recording used for DNN optimization, with no out-of-sample or repeated-trial validation.","rationale":"The reader's weakest assumption correctly flags the absence of identifiability guarantees around Eq. (2), Eq. (6), and the assumptions in Section III. I agree that this is a real structural gap. But I would sharpen it: the empirical comparison to ground-truth HRIRs in both simulation and experiment already provides direct evidence that the method recovered true HRIRs in the tested instances, so the underdetermination concern is not fatal by itself. The more load-bearing weakness is that all quantitative evidence is in-sample: the DNN is optimized on the same recording used to compute NM and LSD, with one noise realization and one experimental measurement and no repeated trials or error bars. The paper's ablations, especially the N/10 sequence updating failure, show that the training scheme is delicate, and a 53-million-parameter network optimized by 300 epochs of BPTT can plausibly fit recording-specific structure. A held-out or cross-noise evaluation would settle whether the identified HRIRs generalize. This concern does not overturn the reader's CONDITIONAL verdict; it reinforces it. The available source code and the ground-truth-based metrics are real supporting evidence, but they do not remove the need for repeated-trial and out-of-sample validation. I therefore recommend no change to the reader's verdict.","tokens_in":20437,"tokens_out":8653,"duration_ms":91032,"concrete_test":"Repeat the full simulation protocol with 10 independent noise realizations while keeping the ground-truth HRIR sequence fixed, and report mean and standard deviation of NM and LSD. In addition, for each noise seed, train the DNN on that seed, freeze the DNN parameters, and evaluate the identified HRIRs on a second, independently generated recording with the same ground-truth HRIRs but a different noise draw. If the cross-seed NM degrades by more than about 2 dB relative to the in-sample NM, or if the across-seed standard deviation of the in-sample NM exceeds about 1 dB, the reported single-run numbers do not establish that the method identifies HRIR dynamics rather than fitting the specific recording.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II explicitly concedes that Eq. (2) is underdetermined at each time index: one scalar observation y(n) versus S*K unknown HRIR coefficients. Section III adds only informal assumptions of slow variation and spatial continuity, and Section IV never converts those assumptions into an identifiability or convergence argument for the DNN update rule of Eq. (6) with training loss Eq. (8). This is a genuine gap. However, the paper's reported NM and LSD are computed against ground-truth HRIRs, so those numbers are direct evidence that the method recovered the true impulse responses in the tested instance, not merely that it fit the output signal. The load-bearing problem is therefore narrower and more specific: the DNN is optimized on the exact recording used for evaluation, with one noise realization in simulation and one measurement in experiment, and no error bars are reported. With 53 million parameters and 300 epochs of BPTT over 17,640-step segments, the optimization has ample capacity to absorb recording-specific artifacts such as a particular noise draw or a particular reflection pattern. If that occurred, the reported NM and LSD would be optimistic and would not transfer to a second recording of the same subject under the same rotation protocol. The central claim, that the method accurately identifies time-varying HRIRs in general, requires evidence that the identified h generalizes beyond the particular recording used to train the update rule.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-neural-network-based method for identifying time-varying head-related impulse responses from a continuously rotating speaker array. The method replaces the fixed step-size update of an adaptive filter with a GRU-plus-FC network that takes the gradient of the instantaneous squared error, the excitation-signal power, and a hidden state as inputs, and outputs an additive update to the HRIR vector. A learnable normalization and a whole-sequence updating/optimization scheme are introduced, and training is performed without a labeled HRIR dataset by minimizing the logarithm of the mean squared prediction error over the recording. The method is evaluated in simulation with interpolated FABIAN HRIRs at 45°/s and in a KEMAR experiment with a custom nine-speaker arc array, reporting NM of -25.58 dB and -23.21 dB and LSD of 1.74 dB and 1.98 dB, respectively, outperforming NLMS, Kalman filter, JO-NLMS, and Meta-AF, with ITD errors below 12 μs.","tokens_in":20756,"tokens_out":8854,"duration_ms":80019,"significance":"If the results are reproducible, this is a valuable contribution: it offers a route to fast individualized HRIR measurement that avoids static repositioning and remains accurate at high rotation speeds, with objective validation against external ground truth in both simulation and hardware. Strengths of the paper are the direct time-domain/frequency-domain comparison to measured or interpolated HRIRs, the inclusion of an ablation study, public source code, and a real KEMAR experiment. The main limitations are that the evaluation consists of single-point estimates from one recording per condition and that the relation between the implemented segment-wise processing and the claimed whole-sequence optimization is not reconciled. These issues affect the strength of the central generalization claim but are addressable with additional experiments and clarification.","major_comments":[{"comment":"Section V-A reports that the full 176,400-sample sequence was divided into 10 segments of 17,640 samples and processed in parallel on 4 GPUs, and Section VI-A says the same was done for the 170,000-sample experimental recording. This is difficult to reconcile with the ablation in Table 1, where the 'N/10 sequence updating and optimization' condition yields NM of 75.29 dB and LSD of 550.71 dB while the proposed method yields -25.58 dB and 1.74 dB. If the reported results are obtained from 10 independent segments, the training loss in Eq. (8) is minimized over 17,640 samples per segment rather than over the whole sequence, so the 'whole-sequence' scheme is not what is evaluated. Please clarify whether the segments share DNN parameters and hidden state, how segment boundaries are handled, and why the N/10 ablation is catastrophic while the parallel segmentation works.","section":"V-A, VI-A, and Table 1"},{"comment":"The NM and LSD numbers in Tables I and III are point estimates from a single simulated noise realization and a single experimental measurement, respectively. Because the DNN is optimized by minimizing the prediction error in Eq. (8) on the same recording that is later used to compute the metrics, the 53-million-parameter model has ample capacity to memorize recording-specific noise or reflections; the reported values therefore do not yet demonstrate that the identified HRIRs generalize to a second recording of the same subject under the same rotation protocol. I request repeated independent noise draws in simulation and repeated rotations in the experiment, reported as mean ± standard deviation, or an explicit out-of-sample evaluation.","section":"Tables I and III, Eq. (8)"},{"comment":"Section II correctly acknowledges that Eq. (2) is underdetermined at each time index, and Section III adds only informal assumptions of slow variation and spatial continuity. The paper never states under what conditions the update rule in Eq. (6), trained with the loss in Eq. (8), recovers the true time-varying HRIR rather than an arbitrary sequence that reproduces y(n). Since the training objective constrains only the microphone output, the favorable NM/LSD values could reflect implicit regularization rather than a guaranteed identification. Please provide either a formal identifiability/convergence analysis for the DNN update or a sensitivity study (varying initialization, noise level, rotation speed, and excitation signal) that supports the claim of accurate HRIR identification beyond the specific tested instance.","section":"Section II, Eq. (2), Eq. (6), Eq. (8)"}],"minor_comments":[{"comment":"In Section V-A the estimated HRIR length is set to K~=192, but Section V-B states zero-padding matches K~=196 to K=256; please correct the value and the corresponding zero-padding description.","section":"V-A and V-B"},{"comment":"Section IV-B's example uses a sampling rate of 48 kHz, whereas the simulation uses 44.1 kHz and the experiment uses 34 kHz; please unify the notation.","section":"IV-B"},{"comment":"Equation (11) as printed does not show the dependence of the correlation sum on τ; please provide the full cross-correlation formula and the range of τ searched.","section":"Eq. (11)"},{"comment":"The phrase 'sampling period in the millisecond range' is likely a typo; for a 44.1 kHz rate the period is about 22.7 μs, not milliseconds.","section":"Section II"},{"comment":"The notation (θ, φ) in the Fig. 8 titles is used without an explicit definition of azimuth and elevation conventions in the text; please add a sentence defining them.","section":"Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is promising and the empirical demonstration is substantial, but the same-recording evaluation and the segment-versus-whole-sequence inconsistency are important. If the authors can provide repeated-trial statistics and clarify the segmentation, I would support acceptance. The underdeterminedness issue is a correctness-risk concern but not, in itself, a reason for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core idea is genuinely new: instead of training a network on a database of HRIRs, the authors train a gated recurrent update rule on a single binaural recording to track the time-varying HRIR vector, with no external labels. That is a sensible and interesting twist on the adaptive-filter literature. Second, every number in the paper comes from the same recording used for the DNN optimization, with no repeated trials or error bars, so we cannot tell how stable the reported NM and LSD are.\n\nWhat the paper does well: the problem formulation is clear, the method is described in enough detail to reproduce, and the evaluation includes both a simulated FABIAN case and a real KEMAR measurement with a custom speaker array. The ablations are informative and show that the GRU gates and the learnable normalization each contribute meaningfully. The ITD analysis is a nice addition and shows errors below 12 microseconds. Source code is promised at GitHub.\n\nThe soft spots are real but not fatal. The underdetermined nature of Eq. (2) is acknowledged, and the paper falls back on slow-variation and spatial-continuity assumptions. Those are reasonable, but a formal block-identifiability argument, noting that the perfect-sweep excitation makes the system identifiable over a window, would have been better than an informal appeal. The bigger issue is the single-instance evaluation. With 53 million parameters and 300 epochs of backpropagation through time, the update rule could in principle adapt to the specific noise draw or the specific reflection pattern of the one experiment. The fact that NM is measured against external ground truth is reassuring, but we do not know how much the result would scatter across different noise realizations or different subjects. A few repeated simulation runs with different noise seeds would substantially raise confidence.\n\nWho should read this: anyone working on fast HRTF measurement, adaptive filtering for time-varying acoustics, or spatial audio capture. It is a solid engineering contribution, not a fundamental advance, and the claims are scoped accordingly. I would send it to peer review. The likely outcome is major revision with requests for error bars, multiple trials, and ideally a second subject or a second recording to show the method is not overfitting to a single instance.","headline":"A genuinely new DNN-based HRIR identification method with solid but single-trial evidence; the per-recording training scheme deserves closer scrutiny.","tokens_in":21244,"tokens_out":4353,"would_cite":true,"duration_ms":39486,"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":"This paper claims that a sequence-to-sequence DNN can identify time-varying head-related impulse responses from a continuously rotating speaker array, achieving normalized misalignment of -25.58 dB in simulation and -23.21 dB in…","keywords":["head-related impulse response","HRIR identification","dynamic HRTF measurement","rotating speaker array","sequence-to-sequence learning","gated recurrent unit","adaptive system identification","learnable normalization"],"falsifier":"Re-run the whole-sequence update on the same binaural recording from different random initializations of the DNN weights; if the identified HRIR sequences differ by more than the reported normalized-misalignment or spectral-distortion margin, then the underdetermined system is being settled by hidden regularization rather than by the data. Equally direct: hold out one azimuth segment of the recording, optimize only on the rest, and ask whether the held-out HRIRs are recovered; failure would mean the identification is fitting the recording instead of discovering a transferable head response.","tokens_in":20257,"feed_emoji":"🎧","tokens_out":12463,"duration_ms":99854,"temperature":0.7,"pith_summary":"The paper aims to make individualized head-related impulse response (HRIR) measurement much faster by letting the speaker array rotate continuously while all speakers play simultaneously, then recovering the time-varying HRIRs from the single microphone output. It claims that a DNN built from gated recurrent units and fully connected layers can do this by replacing the fixed update rule of an adaptive filter: at each time step the network reads the gradient of the instantaneous squared error, the excitation power, and a hidden state, and outputs a coefficient update for the HRIR vector. Because the update rule is optimized on the measurement itself, the method does not need a training dataset of impulse responses, unlike existing DNN-based adaptive filters. The paper reports that this scheme outperforms NLMS, Kalman filtering, JO-NLMS, and Meta-AF, with normalized misalignment of -25.58 dB in simulation and -23.21 dB in experiment, and preserves interaural time differences below the just-noticeable limit.","feed_headline":"Neural net extracts ear sound filters from a spinning speaker rig","feed_subtitle":"No training dataset needed; it beats adaptive-filter baselines and keeps localization cues under 12 microseconds.","key_machinery":"The load-bearing object is the DNN update rule in Eq. (6), $\\tilde{\\mathbf{h}}_{n+1,\\mathrm{ele}} = \\tilde{\\mathbf{h}}_{n,\\mathrm{ele}} + g_\\phi\\{\\nabla \\mathrm{ISE}(n), \\mathbf{x}_{n,\\mathrm{ele}}\\mathbf{x}_{n,\\mathrm{ele}}^\\top, \\mathbf{c}_n\\}$, in which a gated recurrent network turns the gradient of the instantaneous squared error into a coefficient update for the whole HRIR vector. The reset and update gates of the gated recurrent unit carry state across long sequences, the three-layer fully connected block models the nonlinear reflection and diffraction transitions of HRIRs, and the learnable normalization, an input derived from the excitation power, controls the update scale adaptively. The accompanying whole-sequence updating and optimization scheme alternates between rolling the HRIR estimate forward over all time steps with the network frozen and then optimizing the network with backpropagation through time on the accumulated error, which lets the system train without any labeled impulse-response dataset.","core_discovery":"The central claim is that the underdetermined problem of identifying a whole elevation plane of HRIRs from a single scalar microphone signal becomes tractable when the identification rule itself is learned rather than fixed. On a continuously rotating array, the HRIR vector changes smoothly with azimuth, and the proposed network exploits that continuity through a gated recurrent structure: reset and update gates let the update process remember past HRIR states over sequences of hundreds of thousands of samples, while three dense layers convert the instantaneous-error gradient into the next coefficient update. A learnable normalization driven by the speaker excitation power stabilizes the otherwise runaway scale of the gradient. The paper shows that with this update rule, identified HRIRs match ground truth to -25.58 dB normalized misalignment and 1.74 dB log spectral distortion in simulation, and -23.21 dB and 1.98 dB in experiment, and that localization cues, including interaural time differences below 12 microseconds and spectral peaks and notches in the 4-16 kHz range, survive the identification.","pith_inferences":["Editorial inference: the same DNN-in-the-loop coefficient update is not ear-specific, so it could transfer to other time-varying acoustic system identification problems, such as moving sources or changing room responses, wherever one scalar output must resolve many coefficients.","Editorial inference: because the underlying equation is underdetermined, the reported accuracy probably reflects the network's implicit bias toward smooth HRIR sequences as much as the data; a comparison against a smoothness-constrained solver with the same priors would separate identification from regularization.","Editorial inference: the per-measurement training cost, multi-hour and multi-GPU for one sequence, is the practical barrier; amortizing or warm-starting the optimization across subjects or angles is a natural next step that the paper leaves open."],"forward_implications":["A full spherical HRIR set for an individual can in principle be captured in a few seconds of continuous rotation rather than minutes of stepped repositioning.","The method removes the need for a pre-measured corpus of impulse responses for DNN-based identification, so it can adapt to a new subject or a new measurement rig from the single recording itself.","Perceptually relevant localization cues survive the fast-rotation identification: simulation interaural time differences stay below 12 microseconds, and spectral notches and peaks in the 4-16 kHz band are reconstructed closely enough to support azimuth and elevation cues.","At 45 degrees per second the proposed method beats analytic adaptive filters by more than 7 dB in normalized misalignment in simulation, showing the neural update rule tracks faster rotations than the analytical models tolerate."],"supporting_citations":[{"why":"Models the binaural output as a linear time-varying system, giving the underlying equation the method solves.","marker":"[33]"},{"why":"Supplies the perfect-sweep excitation whose circular shifts decorrelate the speakers, enabling simultaneous excitation.","marker":"[35]"},{"why":"Provides the measured HRIR database used as ground truth for the simulation and for interpolated continuous-rotation data.","marker":"[30]"},{"why":"Defines the NLMS baseline whose normalized gradient update the proposed DNN update rule extends.","marker":"[24]"},{"why":"Provides the Kalman-filter baseline and the prior finding that a single linear gain fails at fast rotation, motivating the nonlinear update.","marker":"[29]"},{"why":"Defines the JO-NLMS adaptive-step-size baseline used for comparison.","marker":"[45]"},{"why":"Defines the Meta-AF DNN-based adaptive-filter baseline that requires a training dataset, which the proposed method avoids.","marker":"[44]"},{"why":"Supplies the inverse-distance-weighting interpolation used to build continuous-rotation ground truth from discrete azimuth measurements.","marker":"[48]"},{"why":"Provides the cross-subject HRTF dataset used to train the Meta-AF comparison baseline.","marker":"[37]"}],"fun_headline_variants":["DNN beats analytic models for spinning-speaker HRIR capture","No dataset needed: DNN learns HRIRs from spinning speakers","Gated neural net turns spinning speaker into HRIR scanner","7 dB improvement: DNN extracts HRIRs from rotating speaker","Rotating speaker array plus DNN yields accurate HRIRs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single microphone output per sample, which mathematically leaves the HRIR coefficients underdetermined, still fixes the true HRIR sequence once the DNN's smoothness and continuity priors are applied, so the method's accuracy could be an effect of those priors rather than of information in the recording.","fun_headline_variants_meta":{"raw":{"variants":["DNN beats analytic models for spinning-speaker HRIR capture","No dataset needed: DNN learns HRIRs from spinning speakers","Gated neural net turns spinning speaker into HRIR scanner","7 dB improvement: DNN extracts HRIRs from rotating speaker","Rotating speaker array plus DNN yields accurate HRIRs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001281,"raw_usage":{"total_tokens":5273,"prompt_tokens":1022,"completion_tokens":4251,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":4164}},"tokens_in":638,"tokens_out":4251,"duration_ms":28445,"temperature":1.0,"reasoning_tokens":4164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:48.562529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the whole-sequence update on the same binaural recording from different random initializations of the DNN weights; if the identified HRIR sequences differ by more than the reported normalized-misalignment or spectral-distortion margin, then the underdetermined system is being settled by hidden regularization rather than by the data. Equally direct: hold out one azimuth segment of the recording, optimize only on the rest, and ask whether the held-out HRIRs are recovered; failure would mean the identification is fitting the recording instead of discovering a transferable head response.","supporting_citations":[{"cited_title":"Introduction to digital speech processing,","cited_arxiv_id":null,"evidence_quote":"Models the binaural output as a linear time-varying system, giving the underlying equation the method solves."},{"cited_title":"Perfect-sweep NLMS for time-variant acoustic system identification,","cited_arxiv_id":null,"evidence_quote":"Supplies the perfect-sweep excitation whose circular shifts decorrelate the speakers, enabling simultaneous excitation."},{"cited_title":"A high -resolution head -related transfer function database including different orientations of head above the torso,","cited_arxiv_id":null,"evidence_quote":"Provides the measured HRIR database used as ground truth for the simulation and for interpolated continuous-rotation data."},{"cited_title":"Implementing continuous HRTF measurement in near-field,","cited_arxiv_id":null,"evidence_quote":"Defines the NLMS baseline whose normalized gradient update the proposed DNN update rule extends."},{"cited_title":"Towards faster continuous multi -channel HRTF measurements based on learning system models,","cited_arxiv_id":null,"evidence_quote":"Provides the Kalman-filter baseline and the prior finding that a single linear gain fails at fast rotation, motivating the nonlinear update."},{"cited_title":"An optimized NLMS algorithm for system identification,","cited_arxiv_id":null,"evidence_quote":"Defines the JO-NLMS adaptive-step-size baseline used for comparison."},{"cited_title":"Meta-AF: Meta-learning for adaptive filters,","cited_arxiv_id":null,"evidence_quote":"Defines the Meta-AF DNN-based adaptive-filter baseline that requires a training dataset, which the proposed method avoids."},{"cited_title":"Audibility and interpolation of head -above-torso orientation in binaural technology,","cited_arxiv_id":null,"evidence_quote":"Supplies the inverse-distance-weighting interpolation used to build continuous-rotation ground truth from discrete azimuth measurements."},{"cited_title":"A cross -evaluated database of measured and simulated HRTFs includ ing 3D head meshes, anthropometric features, and headphone impulse responses,","cited_arxiv_id":null,"evidence_quote":"Provides the cross-subject HRTF dataset used to train the Meta-AF comparison baseline."}],"review_version":1}