{"id":"c134a814-9e50-42ea-8bcc-e1c3823368c2","arxiv_id":"2507.20530","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-channel binaural sound event localization and detection system using hand-crafted HRTF cues achieves 4.4 degree localization error on a synthetic benchmark.","lead":"This paper proposes a system that helps robots find and identify sounds using just two microphones, like human ears. It creates a synthetic sound dataset and a custom feature representation, reporting very accurate results on that dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline SELD numbers are measured only on the same 48-direction HRTF grid used for training; the paper never tests off-grid or real binaural input, so the 4.4° error may reflect discrete grid recognition rather than continuous spatial hearing.","rationale":"The paper's internal experiments consistently show monotonic gains from adding each BTFF sub-feature, and the model itself is compact and plausible. The claim I cannot certify is the practical, humanoid-robot-facing one. All reported numbers come from a synthetic set whose directions are exactly the 48 HRTF nodes used at training time; no off-grid, cross-listener, or real-recording evaluation exists. The 4.4° error is far below the 30° grid spacing, so it is compatible with the model having learned to pick the nearest grid node rather than to interpolate spatial cues. The best-of-ten reporting in Table VII further inflates the headline. These are not internal contradictions; they are missing evidence of generalization. A single off-grid test set from the same HRTF database would settle whether the method interpolates spatial cues or merely recognizes training directions. Since the reader already marked the paper conditional on this issue, my read does not change the verdict.","tokens_in":19784,"tokens_out":6133,"duration_ms":73746,"concrete_test":"Using the same Binaural Set synthesis pipeline, construct an off-grid test set with the same 12 event classes and KAIST HRTFs but with directions not in the training grid (e.g., azimuths 15°, 45°, 75°, 105°, 135°, 165° and elevations -15°, +15°, +45°), keeping the same event-pool split. Evaluate the trained BiSELDnet on this set and compare LE_CD and SELD error against the reported Test Set numbers. If LE_CD rises from 4.4° to more than 15°, or SELD error rises from 0.110 to more than 0.2, the low error is grid recognition rather than continuous spatial hearing. Also report mean and standard deviation over the ten training runs instead of only the best run, to assess whether the headline result is an outlier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that BTFF plus BiSELDnet achieves human-like 3D binaural localization with 4.4° error—depends on the Binaural Set being a valid proxy for continuous spatial hearing. Section III-C builds all data from 12 azimuths (30° spacing) × 4 elevations (30° spacing) = 48 KAIST HRTF directions, and every split samples from this same fixed grid. The test clips are rendered by convolving the same isolated-event pool with these same HRIRs, with no off-grid directions, no reverberation, no head movement, and no real binaural recordings. Under this protocol, the model solves a 48-way discrete HRTF-pose recognition problem: the reported 4.4° LE_CD is far below the 30° grid spacing and is consistent with near-perfect selection of the correct grid node, not with interpolation between nodes. The abstract and Table VII also quote the best of ten training runs, so the headline is the upper envelope of the runs. Nothing in the paper demonstrates that the learned cue maps transfer even to the same HRTF set at intermediate azimuths/elevations, let alone to another head or ear geometry. Thus the practical, robot-facing claim is unsupported; the internal ablations are plausible but confined to the grid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BiSELD, a binaural (two-channel) variant of sound event localization and detection, together with a synthetic benchmark dataset (the Binaural Set) built by convolving isolated event recordings from NIGENS and DCASE2016 Task 2 with measured HRIRs from the authors' KAIST HRTF database at 12 azimuths and 4 elevations (48 directions). The proposed input representation BTFF concatenates eight channels—left/right mel-spectrograms, left/right velocity maps, an ITD map (phase-derived delay below 1.5 kHz), an ILD map (above 5 kHz), and left/right spectral-cue maps (mel bands above 5 kHz)—and feeds a 763K-parameter CRNN, BiSELDnet, with ACCDOA-style per-class Cartesian DOA outputs. Ablation experiments show consistent gains from each feature group: V-map improves detection, ITD/ILD maps reduce horizontal-plane localization error (LE from 17.3° to 4.2°), and the SC map reduces median-plane localization error (25.2° to 12.2°). The full system is reported to achieve a SELD error of 0.110, F-score of 87.1%, LE_CD of 4.4°, and LR_CD of 92.1% on the test set.","tokens_in":20103,"tokens_out":13712,"duration_ms":134374,"significance":"If the results hold, the work provides a compact, lightweight two-channel SELD pipeline whose feature design is directly grounded in psychoacoustic cues (ITD below 1.5 kHz, ILD above 5 kHz, pinna-related spectral notches), which is a plausible and falsifiable design hypothesis of genuine interest to the humanoid robotics and binaural-audio communities. The internal ablation logic is clean and consistently executed: each feature is added to a fixed mel-spectrogram baseline and evaluated on the sub-task it is designed for, with ten independent training runs and median reporting in Tables IV–VI. The authors are also transparent about the synthetic nature of the benchmark and about their use of the measured KAIST HRTF database. The main weaknesses are the scope of the evaluation—training and testing on the same 48-direction grid, with no off-grid, reverberant, or real binaural signals—the best-of-ten reporting in Table VII, the absence of any external SELD baseline, and several dataset-description inconsistencies. These issues are local and fixable, but they currently limit the practical force of the headline localization claims.","major_comments":[{"comment":"The entire evaluation is confined to the same 48-direction KAIST HRTF grid used for training, with 30° azimuth and 30° elevation spacing; no test sample is rendered from an off-grid direction, a reverberant scene, a second head or ear geometry, or a real binaural recording. The headline LE_CD of 4.4° in Table VII is an order of magnitude finer than the grid spacing and is consistent with near-perfect selection of the correct grid node rather than with interpolation or continuous spatial inference. Because the abstract and title make claims about 'human-like auditory perception' for humanoid robots, the authors should add an off-grid evaluation (for example, directions at 15° offsets in azimuth and elevation from the training grid) and, ideally, a different HRTF set or real binaural data; until then, the localization claims should be explicitly re-scoped to the grid-matched synthetic setting.","section":"III-C and Table VII"},{"comment":"Tables IV–VI report median values over ten training runs, but Table VII and the abstract report the 'best' performance of BiSELDnet with no measure of variance; the headline SELD error of 0.110, F-score of 87.1%, and LE_CD of 4.4° are therefore the upper envelope of the runs and are not statistically comparable with the median-based ablation tables. The authors should report the median and spread (standard deviation or the full range) for the final configuration in Table VII and ensure that the abstract uses the same statistic.","section":"V-B4, Table VII versus Tables IV–VI"},{"comment":"The evaluation contains no comparison with any external SELD or binaural-localization baseline, even though Table I identifies Wilkins et al. [64] as the state of the art for binaural SELD; without at least one matched baseline trained and evaluated under the same protocol on the Binaural Set, the absolute performance figures have no external calibration and the 'effectiveness' claim rests entirely on self-ablations. Please add such a comparison or explicitly limit the claims to the feature-ablation findings.","section":"Section V and Table I"},{"comment":"The dataset description is internally inconsistent: the text states that 'for each sound class, 20 samples were prepared and split into training, validation, and test sets in a 14:3:3 ratio,' but Table III reports 672/144/144 mixtures whose construction requires far more than 20 event clips per class, and the 14:3:3 ratio does not correspond to 672/144/144 (the actual per-class split implied by Table III is 56/12/12). Additionally, NIGENS is described as having 14 classes but Table II lists 15 class names, the 12 classes actually used in BiSELD are never enumerated, and the exact direction sets underlying Test-H and Test-V are not specified; these details must be corrected and documented for the Binaural Set to serve as a reproducible benchmark.","section":"III-C, Tables II and III"}],"minor_comments":[{"comment":"The Conclusion states that BiSELDnet is built with depthwise separable convolutions, but Section IV and Fig. 4 describe only generic convolution, normalization, activation, and pooling modules; the architecture description and the Conclusion should be reconciled.","section":"Conclusion versus Section IV"},{"comment":"The detection rule 'exceeds 0.5v' appears to refer to a magnitude threshold of 0.5 on the output DOA vector, but the notation is undefined; it should be stated explicitly.","section":"Section IV"},{"comment":"Test-H and Test-V are defined only by their sample counts (36 and 12); the exact azimuth and elevation grid points in each subset should be stated so that Tables V and VI are interpretable from the text alone.","section":"Section III-C"},{"comment":"The spectral-notch analysis quotes elevations from −40° to 90°, whereas the Binaural Set covers only −30° to +60°; the two ranges should be harmonized or the discrepancy explained.","section":"Section III-A3"},{"comment":"The paper should state explicitly that polyphony in the Binaural Set is restricted to across-class overlap, since each mixture contains exactly one instance of each class, and should note the corresponding limitation relative to the same-class overlap scenarios addressed by multi-ACCDOA.","section":"Section III-C"},{"comment":"The abstract's '4.4° localization error' should be qualified: Table VI shows that the median-plane localization error remains 12.2° even with the SC map, so the 4.4° LE_CD is dominated by horizontal-plane performance rather than uniform 3D accuracy.","section":"Abstract and Table VI"},{"comment":"Section IV reports only the total parameter count (763,020); layer-wise specifications such as convolution filter counts, GRU units, and pooling sizes are needed to reproduce BiSELDnet from the text.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior artifacts—the KAIST HRTF database [65], the V-map feature [70], the BiSELD task definition [16, 79], and the underlying PhD dissertation—and the benchmark is evaluated only by the authors' own model. These self-citations are legitimate components, but the absence of any third-party baseline or independent evaluation makes the absolute numbers difficult to calibrate. I would ask the editors to require, at revision, at least one externally grounded comparison (for example, a DCASE-style baseline or the Wilkins et al. binaural configuration) and a clear statement of the dataset and code release plan, since a benchmark paper without a release plan has limited archival value. The manuscript also retains a 'REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER' template placeholder, which should be cleaned before any production stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know upfront: this is a decent feature-engineering paper whose central performance claim is currently unverified. The authors extend binaural SELD to elevation (Wilkins et al. only did the horizontal plane), propose an eight-channel BTFF representation that stacks mel-spectrograms, a velocity map, ITD/ILD maps, and spectral-cue maps, and build a synthetic benchmark by convolving isolated events with measured KAIST HRTFs. The ablations are the best part. Adding V-map helps detection, ITD/ILD maps clearly improve horizontal localization, and the SC-map helps elevation. Each improvement is consistent and matches the stated psychoacoustic rationale. I believe those internal comparisons.\n\nThe soft spot is exactly what the stress-test note says. All training and test data come from the same 48-direction grid (12 azimuths × 4 elevations, 30° spacing) and the same HRTF database. No off-grid directions, no reverberation, no real binaural recordings. The 4.4° LE_CD is far below the 30° grid spacing and is essentially what you would get from perfect 48-way classification. That does not demonstrate continuous 3D localization. The abstract and Table VII report the best of ten runs, while the text says ten runs were done for statistical reliability; medians would be more honest. There is also no comparison with an external SELD baseline, and no dataset or code release, which makes the Binaural Set hard to build on. These are not fatal to the paper's existence, but they are load-bearing for the robot-facing claim in the title and abstract.\n\nI would not call this circular in a harmful sense: the feature design is based on known HRTF cues, and the ablations on held-out synthetic samples are legitimate internal evidence. The citation pattern is fine; self-citations are to the KAIST HRTF database and the authors' earlier V-map work, which are standard components.\n\nWho gets value from this? Researchers working on binaural feature representations or lightweight SELD for robotics. They should read it as a well-structured proof-of-concept, not as a demonstrated practical system. It deserves a serious referee, because the task definition and feature set are useful and the experiments are reproducible in principle. But the review should push for off-grid evaluation, preferably with a second HRTF set or real binaural recordings, and for reporting medians across runs instead of the best run. As it stands, I would not cite the headline numbers as evidence of human-like binaural localization.","headline":"The feature design and ablation logic are solid, but the headline 4.4° localization error is measured on the same 48-direction HRTF grid used for training, so it may reflect grid-node recognition rather than continuous spatial hearing.","tokens_in":20590,"tokens_out":1580,"would_cite":true,"duration_ms":20048,"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":"HRTF cue maps let two microphones localize sounds to 4.4 degrees","keywords":["binaural sound event localization and detection","head-related transfer function","interaural time difference","interaural level difference","spectral cues","sound event detection","direction of arrival estimation","humanoid robots"],"falsifier":"Render binaural test clips from HRIRs at directions not seen in training (for example, 15 degree rather than 30 degree azimuth steps, or elevations at 45 degrees), or record real binaural sounds with a manikin head, and rerun BiSELDnet; if localization error jumps well above 4.4 degrees or detection F-score drops, the central generalization claim fails.","tokens_in":19605,"feed_emoji":"🎧","tokens_out":5911,"duration_ms":55579,"temperature":0.7,"pith_summary":"This paper tries to establish that binaural sound event localization and detection (BiSELD), detecting what is sounding and from where using only two ear-like microphones, can be tackled with hand-designed spatial cues rather than four-channel microphone arrays. To test this, the authors build a synthetic Binaural Set by convolving isolated sound events with measured head-related impulse responses at 48 directions, and propose an eight-channel time-frequency input, the Binaural Time-Frequency Feature (BTFF), that encodes interaural time difference, interaural level difference, and high-frequency spectral cues along with mel-spectrograms and velocity maps. A compact CRNN, BiSELDnet, trained on BTFF reports a SELD error of 0.110, an F-score of 87.1%, and a 4.4 degree localization error on the test set. The paper argues this shows a two-channel, HRTF-informed system can approximate human-like spatial hearing for humanoid robots with much lighter hardware than conventional arrays.","feed_headline":"HRTF cue maps let two microphones localize sounds to 4.4 degrees","feed_subtitle":"An eight-channel feature separates ITD, ILD, and spectral cues so a compact CRNN detects 12 event classes in 3D.","key_machinery":"The central object is the Binaural Time-Frequency Feature (BTFF), an eight-channel input map built from a binaural pair: left and right mel-spectrograms; left and right velocity maps (time-differences of the magnitude spectrogram); an ITD-map, computed as $\\frac{1}{\\omega}\\mathrm{Im}[\\ln(P_R/P_L)]$ for bins below 1.5 kHz and projected to the mel scale; an ILD-map, $10\\log_{10}|P_R/P_L|^2$ for bins above 5 kHz, also mel-projected; and left and right SC-maps, mel spectra restricted to bands above 5 kHz. These channels are the mechanism because they parse the HRTF's spatial information into complementary, frequency-segregated cues before the network sees the data: ITD carries low-frequency azimuth, ILD carries high-frequency azimuth plus front-back asymmetry, and SC carries elevation. The network itself, BiSELDnet, is a CRNN with depthwise separable convolutions, bidirectional GRUs, and a fully connected head that outputs one 3D direction vector per event class per frame, using the activity-coupled Cartesian DOA training target.","core_discovery":"On the paper's own terms, the central discovery is that each class of HRTF-derived cue maps to a specific sub-problem of binaural SELD, and encoding them explicitly makes the whole task learnable from two channels. The velocity map improves detection by marking onsets and transients; the ITD-map, derived from the imaginary part of the log spectral ratio below 1.5 kHz without phase unwrapping, and the ILD-map, computed as log-power difference above 5 kHz, jointly resolve azimuth, with ILD's front-back asymmetry disambiguating ITD's symmetry; and the SC-map, mel bands above 5 kHz, supplies elevation-dependent pinna notch cues. With these eight channels, BiSELDnet detects all 12 event classes and outputs a 3D direction vector per class per frame, reaching a 4.4 degree localization error and 92.1% localization recall on the Binaural Set test split. The authors take this as evidence that HRTF-based spatial cues integrated as input features are a viable path to binaural 3D SELD in humanoid robots.","pith_inferences":["Because training and test directions come from the same 48-point HRTF grid, the 4.4 degree error may partly measure interpolation inside that grid; the paper does not establish accuracy on directions between grid points or on real recorded binaural scenes.","A natural next test is to evaluate the same BTFF on continuous azimuths and on real binaural recordings from a manikin head; if accuracy degrades, the cues remain valid but the current benchmark overstates deployable performance.","The feature design suggests a hearing-aid or telepresence variant could work with generic HRTFs per user, provided individual pinna cues are preserved; the SC-map channel is the load-bearing part for elevation and the most likely to need per-listener adaptation.","One could ablate the hand-crafted maps against learned spatial features, for example a network given only the two mel-spectrograms, to quantify how much of the 4.4 degree result comes from the engineered cues versus the CRNN itself."],"forward_implications":["A two-channel binaural front end is enough for joint 3D detection and localization in the synthetic setting, so robots can avoid the size, calibration, and data costs of four-channel arrays.","Explicitly separating ITD, ILD, and spectral cues makes each cue's contribution inspectable: V-map for detection, ITD/ILD for azimuth, and SC for elevation.","The reported 4.4 degree localization error and 92.1% recall imply the model can pick out which of 12 classes is active and point to it with near-grid resolution on the Binaural Set.","The Binaural Set itself, with clean and noisy conditions and separate horizontal and vertical test subsets, gives the community a controlled benchmark for comparing binaural SELD methods."],"supporting_citations":[{"why":"Supplies the CRNN-based SELD framework and the joint detection-localization output formulation this model builds on.","marker":"[28]"},{"why":"Provides the activity-coupled Cartesian DOA representation used as the training target for each event class.","marker":"[47]"},{"why":"Prior binaural SED study showing binaural features help detection, motivating the input design.","marker":"[51]"},{"why":"The main prior binaural SELD comparison, which restricted evaluation to the horizontal plane and is extended here to 3D.","marker":"[64]"},{"why":"Supplies the measured HRTF database used to generate the Binaural Set and the cue analysis of spectral peaks and notches.","marker":"[65]"},{"why":"Psychoacoustic reference establishing ITD, ILD, and spectral cues as the localization mechanisms the features encode.","marker":"[66]"},{"why":"Provides the isolated foreground sound event database used in synthesizing the Binaural Set.","marker":"[76]"},{"why":"Adds real-world sound event recordings for foreground diversity and supplies the DCASE evaluation protocol.","marker":"[77]"},{"why":"Defines the joint SELD metrics, including the SELD error formula used to report the headline results.","marker":"[80]"}],"fun_headline_variants":["Two-mic robot ears hit 4.4° localization with HRTF cue maps","HRTF feature maps give two mics human-like 3D sound localization","Four degree accuracy from binaural cues: BiSELDnet's HRTF maps","Explicit ITD, ILD, and spectral cues sharpen binaural SELD to 4.4°","Two-mic SELD with HRTF cue maps: 87.1% F-score, 4.4° error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that performance measured on synthetic mixtures made from a fixed 48-direction HRTF grid transfers to real binaural hearing; if the grid is not representative of continuous, real-world directions, the reported accuracy is a property of the benchmark rather than of the robot's ears.","fun_headline_variants_meta":{"raw":{"variants":["Two-mic robot ears hit 4.4° localization with HRTF cue maps","HRTF feature maps give two mics human-like 3D sound localization","Four degree accuracy from binaural cues: BiSELDnet's HRTF maps","Explicit ITD, ILD, and spectral cues sharpen binaural SELD to 4.4°","Two-mic SELD with HRTF cue maps: 87.1% F-score, 4.4° error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2833,"prompt_tokens":1076,"completion_tokens":1757,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":1633}},"tokens_in":692,"tokens_out":1757,"duration_ms":13139,"temperature":1.0,"reasoning_tokens":1633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:42:08.501706+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render binaural test clips from HRIRs at directions not seen in training (for example, 15 degree rather than 30 degree azimuth steps, or elevations at 45 degrees), or record real binaural sounds with a manikin head, and rerun BiSELDnet; if localization error jumps well above 4.4 degrees or detection F-score drops, the central generalization claim fails.","supporting_citations":[{"cited_title":"Two vs. four-channel sound event localization and detection,","cited_arxiv_id":null,"evidence_quote":"The main prior binaural SELD comparison, which restricted evaluation to the horizontal plane and is extended here to 3D."},{"cited_title":"ACCDOA: Activity -coupled Cartesian direction of arrival representation for sound event localization and detection,","cited_arxiv_id":null,"evidence_quote":"Provides the activity-coupled Cartesian DOA representation used as the training target for each event class."},{"cited_title":"A report on sound event detection with different binaural features ,","cited_arxiv_id":null,"evidence_quote":"Prior binaural SED study showing binaural features help detection, motivating the input design."},{"cited_title":"HRTF measurement for accurate sound localization cues","cited_arxiv_id":"2203.03166","evidence_quote":"Supplies the measured HRTF database used to generate the Binaural Set and the cue analysis of spectral peaks and notches."},{"cited_title":"Iida, Head-Related Transfer Function and Acoustic Virtual Reality","cited_arxiv_id":null,"evidence_quote":"Psychoacoustic reference establishing ITD, ILD, and spectral cues as the localization mechanisms the features encode."},{"cited_title":"Joint measurement of localization and detection of sound events ,","cited_arxiv_id":null,"evidence_quote":"Defines the joint SELD metrics, including the SELD error formula used to report the headline results."}],"review_version":1}