{"id":"e298616b-dac4-4d54-81bd-fbd1b9471acd","arxiv_id":"2411.13179","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A network trained only on simulated audio, SONNET, estimates microphone time delays on real-world recordings more accurately than GCC-PHAT and improves downstream self-calibration.","lead":"The paper trains SONNET, a neural network, on simulated room audio, and shows it estimates the time delay between two microphones more accurately than the standard GCC-PHAT method on real recordings with known ground truth. The result matters because accurate time delays are the foundation of sound localization, microphone calibration, and beamforming.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulation hyperparameters were ablated on the single real test set (Sec. 6.5), so the 'novel real data' claim needs an independent real-world evaluation before it holds.","rationale":"SONNET's architecture and training are standard and the simulated-data experiments are carefully done. The paper also provides code and a pretrained model, which is good. However, the central claim is a sim-to-real generalization claim, and the only real-world evidence is a single dataset whose simulation configuration was ablated on that same dataset. The ablation in Section 6.5 is the key: it shows that adding movement and directionality improves the model on tdoa_20201016. This is an implicit selection of the training distribution using test labels. Thus the model may be overfit to the acoustic conditions of that particular dataset. The qualitative plots (Fig. 5) and downstream self-calibration (Table 2) are convincing within this dataset, but they do not by themselves establish 'novel real world data' in a general sense. The reader's conditional verdict is appropriate; my concern reinforces it rather than changing it. A second independent dataset is the decisive check. I agree with the reader's weakest assumption, but I sharpen it: the issue is not just that the simulation might not cover all real conditions, but that the simulation was in part chosen to match the specific real test set.","tokens_in":9577,"tokens_out":6653,"duration_ms":66765,"concrete_test":"Evaluate the released SONNET model, with no fine-tuning or hyperparameter changes, on a second independent real-world TDOA dataset with ground truth (e.g., the AV16.3 dataset or a newly recorded array in a different room with different hardware), computing the same inlier ratio at 10 cm used in Fig. 4. If SONNET does not exceed GCC-PHAT by a meaningful margin (e.g., >5 percentage points) on this held-out dataset, the central generalization claim is not established. If a second dataset is unavailable, an internal split of tdoa_20201016 (e.g., held-out experiment types) with re-selection of the simulation augmentations on the training split would be a weaker but feasible alternative.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SONNET, trained only on simulated audio, significantly outperforms GCC-PHAT on novel real-world data. For that claim to hold, the real test data must be genuinely novel with respect to the training design. This condition is not met in the paper: the simulation recipe (cuboid rooms 1–10 m, reflection coefficients 0.05–0.99, subcardioid directivity, moving sources) is not a fixed a priori design but was validated and selected via ablation studies on the very same tdoa_20201016 dataset used for the headline comparison (Sections 6.3 and 6.5, Fig. 7a). The augmentations (movement, directionality) were kept because they improved accuracy on this specific real dataset. Consequently, the reported margin over GCC-PHAT on this dataset may reflect selection on the test distribution, not generalizable sim-to-real transfer. Moreover, the real-data evaluation is based on a single dataset (one recording setup, one microphone array geometry), so 'novel real world data' is an extrapolation from n=1. If the model has latched onto statistics particular to this dataset—microphone hardware, room acoustics, or source types—the advantage could vanish on other real recordings. This is the weakest load-bearing assumption; the rest of the pipeline (regression-via-classification, ResNet, etc.) is internally sound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SONNET, a learned model for time-delay estimation (TDE) from pairs of audio recordings. The model is trained exclusively on a large simulated dataset generated with Pyroomacoustics, using random rectangular rooms, wall reflection coefficients in [0.05, 0.99], subcardioid directivity, and sound sources that are either stationary or follow discretized quadratic Bézier paths. The network uses a Fourier front-end, a convolutional backbone, and a ResNet with regression-via-classification over 1000 one-sample TDOA bins. The authors claim that SONNET, trained only on synthetic data, significantly outperforms GCC-PHAT on novel real-world recordings, and they demonstrate improved downstream self-calibration using SONNET-based TDOA estimates. The experiments include inference speed measurements, simulated noise/reverberation sensitivity tests, an evaluation on the real tdoa_20201016 dataset with ground truth, self-calibration results, and ablations over the simulation augmentations and dataset size.","tokens_in":9823,"tokens_out":6492,"duration_ms":67247,"significance":"If the central claim holds, the paper offers a practical and immediately usable TDE model that improves over a classical baseline without retraining on real data. The strengths are the release of a pretrained model, the use of an external real-world dataset with ground truth, the downstream self-calibration evaluation, and the ablation studies that identify which simulation choices matter. However, the generality of the result is currently supported by only one real-world dataset, and the simulation design choices were, in part, selected using that same dataset. These issues do not invalidate the pipeline, but they do require additional evidence or more cautious claims before the paper can be accepted.","major_comments":[{"comment":"The ablation study that motivates the two central simulation augmentations (source movement and directionality) is evaluated on tdoa_20201016, the same dataset used for the headline real-data results in Section 6.3 and Fig. 4. The full SONNET configuration was therefore retained, at least partly, because it improves accuracy on this exact dataset. As a result, the dataset cannot be considered 'novel' with respect to all design choices, and the reported margin over GCC-PHAT may partly reflect selection on the test distribution. Please provide an additional real-world evaluation on a dataset that played no role in any ablation or design decision, or use a nested holdout procedure in which the final model is evaluated only on data not used for any model selection.","section":"Section 6.5, Fig. 7a"},{"comment":"The real-world evidence rests entirely on a single dataset, tdoa_20201016, which appears to involve one recording environment and one microphone configuration. The abstract's claim of 'novel real world data' and Section 1.3's claim of generalization to real data extrapolate from this single dataset. Please add at least one independent real-world TDE dataset with ground truth, or explicitly limit the claims to the tested dataset and describe what properties of the dataset are needed for the result to transfer.","section":"Section 6.3 and abstract"},{"comment":"The paper reports inlier-ratio curves and RMS errors without error bars, confidence intervals, or significance tests. Since the abstract says SONNET 'significantly outperforms' GCC-PHAT, the word 'significantly' needs statistical support. Furthermore, the 384,648 window pairs come from a small number of rooms and recordings, so pairs are not independent; any confidence intervals or tests should account for clustering by room or recording, or the effective sample size and the significance claim will be unclear.","section":"Figures 3, 4, 7 and Table 2"},{"comment":"The contribution statement says SONNET 'outperforms state of the art methods for TDE', but the experiments compare only with GCC-PHAT. The related work lists several learned TDE methods (refs. 9-11, 25), so a comparison with at least one learned baseline trained on similar simulated data is needed to support the 'state of the art' claim. Alternatively, the claim should be rephrased to say that SONNET outperforms GCC-PHAT, which is what the experiments actually show.","section":"Section 1.3 and Section 6"}],"minor_comments":[{"comment":"The notation hi(t, j/k) in Eq. (4) is ambiguous; it should be hi(t, t_j) with t_j = jT/k. In addition, the surrounding sentence about 'simulating part x(t), i-1 k T < t < i k T' is garbled by the formatting and should be rewritten with clear parentheses.","section":"Section 3.1, Eq. (4)"},{"comment":"The simulated evaluation in Fig. 3 uses audio from tdoa_20201016 played through simulated rooms. This is a reasonable stress test, but it should be described as testing robustness to those source types in synthetic reverberation, not as a measure of generalization to novel real-world reverberation.","section":"Section 6.2"},{"comment":"The caption states that 'GCC-PHAT is very robust against white noise' without explaining that this refers to the flatness of the curve; please clarify the intended meaning.","section":"Fig. 3 caption"},{"comment":"The legend entries 'SONNET (d)', 'SONNET (m)', and 'SONNET (d+m)' are not defined in the caption. Please define the abbreviations in the caption or refer explicitly to the text where they are introduced.","section":"Fig. 7a caption"},{"comment":"Table 1 would benefit from stating the CPU model and whether the reported times include preprocessing or only the network forward pass; the text gives the CPU model but the caption does not.","section":"Table 1"},{"comment":"The 12 million training pairs are generated from only 10,000 rooms, so pairs from the same room share the room impulse response and are not independent. This should be acknowledged when discussing dataset scale and generalization.","section":"Section 5"},{"comment":"Please provide a brief description of the tdoa_20201016 recording setup (room size, microphone layout, source types) or a more specific reference to where it is described, since the paper's main evaluation depends on this dataset.","section":"Section 6.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with a clear practical contribution, but the central claim as stated in the abstract is stronger than the evidence. The main risk is that the single real dataset was used both for ablation-driven design choices and for the final evaluation, which weakens the 'novel real world data' claim. I believe the paper can be revised by adding an independent real-world evaluation or by carefully rephrasing the claims, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SONNET reports a genuine sim-to-real result: a TDE model trained on 12 million simulated pairs, with moving sources and directional microphones, beats GCC-PHAT on real recordings from tdoa_20201016 and improves downstream self-calibration. The story is clean, the code and model are public, and the paper is honest about what previous learning-based work did not claim. Credit where due: the simulation scale is substantial, the ablation studies are informative, and evaluating on real ground-truth data plus a downstream task is the right way to make the case.\n\nThe main soft spot is the \"novel real world data\" phrasing. The simulation choices—movement, directionality, dataset size—were validated by ablations on tdoa_20201016, the same dataset used for the main comparison. So the reported margin over GCC-PHAT reflects in part selection on that test distribution. That doesn't sink the result, but it means the headline claim should be read as \"generalizes to this dataset,\" not \"generalizes to arbitrary real recordings.\" A second independent real recording setup is needed before calling it general.\n\nTwo smaller issues. One: no error bars or significance tests on the real-data comparisons; the margin looks large, but we don't know its stability. Two: no comparison to other learned TDE baselines. The paper cites prior learned methods as not claiming real-data generalization, but a reader still wants to see how SONNET compares to a trained-on-simulated baseline from that literature, if one exists.\n\nThe math and pipeline are internally sound. The moving-source simulation via piecewise stationary convolution is a reasonable approximation and is described well. I don't see a load-bearing flaw. The single-dataset evaluation is the main limitation, and it's a fixable one.\n\nWho this is for: people working on TDE, microphone-array self-calibration, or sim-to-real audio. It deserves a serious referee—send it to peer review. I'd recommend requiring an independent real-world evaluation or at least a much more cautious claim about novelty, plus error bars, and ideally a comparison to a prior learned method. If those are added, it becomes a solid contribution.","headline":"A credible sim-to-real TDE result, but the 'novel real data' claim is softened by model selection on the single real test set.","tokens_in":10389,"tokens_out":2755,"would_cite":true,"duration_ms":27255,"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":"SONNET, a neural network trained entirely on simulated room audio, estimates inter-microphone time differences more accurately than the standard GCC-PHAT method on novel real recordings, and the better estimates let a downstream…","keywords":["time delay estimation","TDOA","GCC-PHAT","sim-to-real transfer","synthetic audio","self-calibration","microphone arrays","deep learning"],"falsifier":"Take SONNET without retraining and evaluate it on a second real-world TDOA dataset with ground truth, recorded in different rooms, with different microphone hardware and different sound content; if its inlier ratio at a 10 cm threshold does not beat GCC-PHAT, the generalization claim is falsified. A cheaper check is to retrain the model with the movement or directionality augmentations removed and observe whether the real-data performance gap the paper reports disappears.","tokens_in":9345,"feed_emoji":"🎤","tokens_out":8873,"duration_ms":76964,"temperature":0.7,"pith_summary":"The paper tries to establish that a neural network can estimate the time difference of arrival (TDOA) between two microphones better than the classical Generalized Cross-Correlation Phase Transform (GCC-PHAT) method, even when trained only on simulated audio and evaluated on real recordings it has never seen. It builds a large synthetic training set covering 10,000 rooms with varied sizes and reverberation, moving sound sources, and directional microphones, and uses it to train SONNET, a real-time model. On the real tdoa_20201016 dataset, SONNET's inlier rate at a 10 cm threshold is higher than GCC-PHAT's, and its TDOA values let a published self-calibration system converge on all test experiments with a median receiver error of 0.05 m, compared to 0.59 m for GCC-PHAT. If true, this shows that simulation can replace the difficult step of collecting ground-truth TDOA labels for training.","feed_headline":"Simulated audio beats the standard method on real recordings","feed_subtitle":"Trained on 12 million simulated pairs, the model improves time-difference estimates and calibration on real recordings.","key_machinery":"The central object is the SONNET model: two recordings are transformed with a fast Fourier transform, frequencies above 4800 Hz are discarded, and the real and imaginary parts are fed as separate channels through three 1D convolutional layers, followed by stacked residual blocks with skip connections, ending in a linear layer over 1000 TDOA classes. The model is trained with cross-entropy loss and label smoothing on 12 million pairs of simulated recordings, taken from 50 choose 2 pairs in each of 10,000 rooms. The simulation machinery matters as much as the network: moving sources are approximated by splitting each source signal into parts and simulating each part as a stationary source at successive points along a path, and directional microphones use a subcardioid sensitivity pattern. This combination is what the paper argues closes the gap between synthetic training data and real audio.","core_discovery":"On the paper's own terms, the discovery is that the sim-to-real gap for audio time-delay estimation is closable by scaling up and enriching the simulation rather than by adding more network complexity. The training set is generated with an image-source room simulator: rectangular rooms with side lengths sampled uniformly from 1 to 10 meters, wall reflection coefficients from 0.05 to 0.99, 50 microphones per room, and sources that are either stationary or moving along smooth quadratic paths at speeds up to 5 m/s, modeled as discretized jumps. Each source is given a subcardioid directional sensitivity pattern. The trained model outputs logits over 1000 classes, each corresponding to a one-sample TDOA bin at 16 kHz, matching GCC-PHAT's resolution. On real data, SONNET achieves a higher inlier ratio at every tested threshold, and the ablation studies show that source movement contributes more than directionality and that larger training sets generalize better.","pith_inferences":["If the sim-to-real transfer holds beyond this one dataset, the same large-scale simulation strategy could be applied to other signal modalities such as radio or ultrasound, where ground-truth TDOA labels are equally scarce.","A natural next test is a second independent real-world corpus with different rooms and hardware; the current evidence rests on a single dataset, so that experiment would directly measure how representative that dataset is.","Because the network outputs a full distribution over delay classes, it could be used to flag uncertain or outlier estimates before feeding downstream solvers, an extension the paper does not explore.","The ablation result that movement matters more than directionality suggests that simulating dynamic sources is the key ingredient; testing even richer motion models, such as continuous Doppler shifts, is a concrete next step."],"forward_implications":["A ready-to-use, real-time TDE model exists for ad-hoc microphone pairs: SONNET runs in about 0.94 ms per pair on a CPU and 0.022 ms on a GPU, roughly four times slower than GCC-PHAT but still real-time.","Self-calibration of microphone arrays becomes far more reliable: with SONNET's TDOA estimates the calibration system converges on all tested recordings, whereas with GCC-PHAT several experiments fail or give errors above 1 m.","The synthetic-data recipe can be reused to train estimators for problems where ground-truth TDOA is hard to collect, such as multiple simultaneous sources, multipath components, or more than two microphones.","The model's one-sample-bin output at 16 kHz means the accuracy gain is not achieved by coarser quantization than the baseline."],"supporting_citations":[{"why":"image-source method that generates the acoustic impulse responses used to build the simulated training rooms","marker":"[1]"},{"why":"prior learning-based TDE model that supplies the regression-via-classification setup and the baseline the authors extend","marker":"[2]"},{"why":"the generalized cross-correlation phase transform method that is the classical baseline SONNET is compared against","marker":"[15]"},{"why":"fast stratified self-calibration system that consumes the TDOA estimates in the downstream experiment","marker":"[17]"},{"why":"room acoustics simulator used to compute impulse responses for the synthetic dataset","marker":"[22]"},{"why":"open corpus of music, speech, and noise that provides the source audio signals in the simulations","marker":"[23]"},{"why":"real dataset with ground-truth TDOA values that provides the novel real-world evaluation and the self-calibration testbed","marker":"[30]"}],"fun_headline_variants":["Synthetic audio training beats classical TDOA on real data","SONNET: neural time-delay estimator beats GCC-PHAT on real audio","AI time-delay estimator trained on simulation surpasses GCC-PHAT","12M simulated pairs improve real-world time delay estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated rooms, with randomly sampled sizes, reflection coefficients, moving sources, and directional microphones, faithfully represent the acoustic conditions of real recordings such as the tdoa_20201016 dataset, and that this one dataset stands for novel real-world data generally.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic audio training beats classical TDOA on real data","SONNET: neural time-delay estimator beats GCC-PHAT on real audio","AI time-delay estimator trained on simulation surpasses GCC-PHAT","12M simulated pairs improve real-world time delay estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001128,"raw_usage":{"total_tokens":4685,"prompt_tokens":938,"completion_tokens":3747,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":3672}},"tokens_in":554,"tokens_out":3747,"duration_ms":27498,"temperature":1.0,"reasoning_tokens":3672,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:44:31.489230+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take SONNET without retraining and evaluate it on a second real-world TDOA dataset with ground truth, recorded in different rooms, with different microphone hardware and different sound content; if its inlier ratio at a 10 cm threshold does not beat GCC-PHAT, the generalization claim is falsified. A cheaper check is to retrain the model with the movement or directionality augmentations removed and observe whether the real-data performance gap the paper reports disappears.","supporting_citations":[{"cited_title":"The Journal of the Acoustical Society of America65(4), 943–950 (1979)","cited_arxiv_id":null,"evidence_quote":"image-source method that generates the acoustic impulse responses used to build the simulated training rooms"},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"prior learning-based TDE model that supplies the regression-via-classification setup and the baseline the authors extend"},{"cited_title":"In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2021)","cited_arxiv_id":null,"evidence_quote":"fast stratified self-calibration system that consumes the TDOA estimates in the downstream experiment"},{"cited_title":"In: 2018 IEEE interna- tional conference on acoustics, speech and signal processing (ICASSP)","cited_arxiv_id":null,"evidence_quote":"room acoustics simulator used to compute impulse responses for the synthetic dataset"}],"review_version":1}