{"id":"d8198934-699d-4a49-a88d-49fa9713b9f7","arxiv_id":"2411.11222","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A learned audio model infers liquid level, container size, flow rate, and fill time from the changing pitch of pouring water, trained with synthetic sounds and video co-supervision.","lead":"This paper shows that a computer can listen to the sound of water being poured into a container and estimate how full the container is, how large it is, and how fast it is being filled. The authors train a pitch-detecting neural network using simulated pouring sounds and real videos, then test it on new containers and YouTube clips.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline air-column accuracy (0.60 cm) may be partially circular: the audio net is co-supervised to match video pseudo-labels, and Table 2's ground truth for l(t) appears to come from the same pseudo-label pipeline; independent liquid-level ground truth is needed.","rationale":"The reader's designated weakest assumption is that the observed pitch is a clean single-valued axial resonance; that is a real scope limitation, but the paper discloses it explicitly (Sec. 6.4, Fig. 15) and the core results are scoped to cylinder-like containers where the physics is standard. The more load-bearing problem is the evaluative circularity in the headline dynamic metric: the co-supervised audio model is trained to match video-derived pseudo-labels, and the ground truth for the l(t) evaluation in Table 2 appears to be those same pseudo-labels rather than an independent measurement. This is not stated in the paper, and it directly affects the paper's most cited number (0.60 cm) and the 'metric ruler' framing. The static properties and flow-rate/time-to-fill evaluations are checked against manual measurements or fill timing, so the physics pipeline retains independent support, but the air-column-length result needs a non-circular check. The proposed test would settle whether the reported accuracy reflects true physical inference or teacher-fitting. Since the reader already issued a CONDITIONAL verdict partly on these grounds, our concern does not change the verdict, but it sharpens the condition: release independent l(t) ground truth and re-report Table 2. Agreement with the reader is partial because the reader flagged circularity as a secondary issue, whereas we regard it as the single most load-bearing concern for the central claim.","tokens_in":21937,"tokens_out":7136,"duration_ms":70982,"concrete_test":"Re-evaluate Table 2 on a subset of Test set I using independently measured liquid-level ground truth: e.g., manual frame-by-frame annotation of the water surface in the videos, or a flow-meter/scale recording synchronized with the audio. Compare co-supervised and audio-only MAE for l(t) against these independent labels. If the co-supervised MAE remains near 0.60 cm, the circularity concern is resolved; if it rises toward or above the audio-only 0.78 cm, the reported co-supervision gain is an artifact of shared supervision targets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest quantitative claim is the air-column-length MAE of 0.60 cm on Test set I (Table 2), which anchors the 'metric ruler' statement (Sec. 3.1) and the 'human-like capabilities' claim (Sec. 1). The audio network is fine-tuned by minimizing MSE against l_px(t) + beta*R_px predicted by the video network (Eq. 10, Sec. 4.3). The video network is itself trained on pseudo-labels obtained from temporal-difference heatmaps and a RANSAC polynomial fit (Appendix A.3). The paper does not state the source of ground-truth l(t) used in Table 2. If it is the same video pseudo-label pipeline, then the co-supervised model is evaluated against its own training target: shared noise and systematic biases (e.g., liquid-surface detection lag, RANSAC smoothing) are not penalized, so the reported 0.60 cm error may measure agreement with the video teacher rather than physical truth. The audio-only variant (0.78 cm) is not fine-tuned on these labels, so the apparent gain from co-supervision could be an artifact of fitting the teacher. This circularity does not invalidate the static-property evaluations (H, R, Q, tau are checked against manual measurements and fill timing), which gives the physics pipeline independent support, but it undermines the headline dynamic metric that the strongest claim relies on. The paper's own failure cases (Sec. 6.4) show the clean single-mode axial-resonance assumption is not universal, yet that scope limitation is honestly disclosed; the evaluation circularity is not disclosed and is therefore the more load-bearing concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper claims that, from only audio of liquid pouring, physical properties of the container-liquid system can be recovered: air-column length l(t), container height H, radius R, volume flow rate Q(t), and time-to-fill τ. The authors derive these properties from the axial-resonance relation λ(t)=4(l(t)+βR) in Section 3, train a wav2vec2-based wavelength-prediction network using simulated pouring sounds and visual co-supervision from a DINO-based video teacher in Section 4, introduce a new 805-video dataset in Section 5, and report quantitative results for these properties, plus shape classification and liquid-mass estimation on an external dataset in Section 6. The central theoretical derivation is mathematically straightforward, and the main empirical concern is the provenance of the air-column ground truth used for the headline result.","tokens_in":22177,"tokens_out":9375,"duration_ms":89320,"significance":"If the empirical claims are taken at face value, the paper makes a meaningful advance: it gives a simple and largely parameter-free physical mapping from pitch to metric properties, demonstrates that a learned pitch detector can outperform classical pitch trackers by a large margin, and introduces a valuable dataset for audio-visual physical inference. The independent manual-ruler evaluation of H and R in Table 3 and the external-dataset mass-estimation results are concrete strengths, and the failure cases in Section 6.4 are honestly disclosed. However, the strongest dynamic-property claim (0.60 cm air-column MAE) currently lacks a demonstrated independent ground-truth source, which is essential before the 'audio as metric ruler' and 'human-like capabilities' claims can be accepted.","major_comments":[{"comment":"The manuscript does not state the source of the air-column ground truth used in Table 2. The co-supervised audio model is fine-tuned with the MSE objective in Eq. (10), where the target is the video network's prediction l_px(t)+βR_px, and the video network is trained on pseudo-labels obtained from temporal-difference heatmaps with a RANSAC polynomial fit (Appendix A.3). If Table 2 evaluates against the same pseudo-label pipeline, the reported 0.60 cm error is not a measurement against physical liquid level but an agreement score with the video teacher, and the apparent gain of co-supervision over the audio-only variant (0.60 vs 0.78 cm) could be largely an artifact of fitting that teacher. This is load-bearing because the 'audio is effectively a metric ruler' claim (Section 3.1) and the claimed 'human-like capabilities' (Section 1) rest on this number. The manual ruler measurements for H and R in Table 3 provide independent support for the static-property chain, but they do not validate the dynamic l(t) curve. Please specify the ground-truth source and, if it is the pseudo-label pipeline, add independent manual or sensor-based liquid-level annotations for at least a subset of videos, reporting both pseudo-label-based and independent errors.","section":"Section 6.1, Table 2; Appendix A.3; Eq. (10)"},{"comment":"The per-video scale factor α is estimated from the ratio between the audio network's own wavelength predictions and the video network's pixel measurements, weighted by RMS energy, and is then fixed while the same audio network is fine-tuned toward Eq. (10). This makes the co-supervision loop partially self-referential: systematic errors in the pre-fine-tune audio model can be absorbed into α and are then not penalized by the objective. The verification in Appendix A.3 (α in [30,80], inverse relation with container size) is only a sanity check. Please provide a sensitivity analysis, e.g., compute α from ground-truth wavelengths on a subset of videos or from an independent metric-to-pixel calibration, and show how the final property errors in Table 3 change. This matters because α is the bridge that converts audio wavelengths into metric quantities used during training.","section":"Section 4.3, Eq. (10)-(11); Appendix A.3"}],"minor_comments":[{"comment":"The split arithmetic is unclear: Table 1 reports 18 train containers/195 videos, 13 Test I containers/54 videos, 19 Test II containers/327 videos, and 25 Test III containers/434 videos, while the text says the totals are 18 + 25 = 43 containers and 195 + 54 + 434 = 683 videos, omitting Test II and the overlap between Test II and Test III; please clarify the unique-container counts and how the 122 remaining videos are defined.","section":"Section 5, Table 1"},{"comment":"The caption lists '(a) Hemispherical container (cup)' and '(b) Bottle-neck container', but the body text describes (a) as a bottleneck case and (b) as a hemispherical case; the caption and text should be made consistent.","section":"Section 6.4, Figure 15"},{"comment":"The column header 'Synthetic ↓' is confusing because the text refers to this model as 'audio-only'; using one consistent name would improve readability.","section":"Table 3"},{"comment":"The rows for time-to-fill use the notation 'τ 1 4 (t)', 'τ 1 2 (t)', and 'τ 3 4 (t)'; please define this notation explicitly, since it is not obvious that these denote the fraction of the original audio given to the model.","section":"Section 6.1, Table 3"},{"comment":"The verification that empirical scale factors are in [30,80] relies on 'generic values' of f, s, and Z, but those values are not stated; please provide the assumed values and units so the range can be reproduced.","section":"Appendix A.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready in its current form because the headline air-column accuracy may be self-referential, but the issue is addressable with additional validation. If the authors provide independent liquid-level ground truth or demonstrate robustness of Table 2 to the ground-truth source, I would support acceptance. The theoretical derivation, dataset, and static-property evaluations are solid, and the failure cases are honestly disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a real contribution, and the physics-based training pipeline is the novel part. The headline number—0.60 cm MAE on air column length—looks partially circular, and that needs to be fixed before I'd trust the dynamic claims.\n\nWhat's genuinely new: the paper doesn't claim new acoustics. The equations are textbook. But they build a working system: a DDSP-based simulator that generates pouring sounds with specified pitch profiles, a wav2vec2 pitch detector pre-trained on those synthetic sounds and fine-tuned with a scale-aware visual co-supervision loss, and a new 805-video dataset. The learned model recovers container height, radius, flow rate, and time-to-fill from audio alone. On Wilson et al.'s weight estimation benchmark, their linear probe beats the supervised baseline on average. That's solid engineering and a useful dataset. The generalization to YouTube videos, while qualitative, is honestly presented.\n\nThe soft spots, in order of importance. First, the air-column evaluation. The video teacher is trained on pseudo-labels from a temporal-difference/RANSAC pipeline, and that same pipeline seems to supply the ground truth for l(t) in Table 2. If that's the case, your co-supervised model is being measured against its own training target. The 0.60 cm error could just be agreement with the video teacher, not physical truth. This doesn't sink the paper—the static properties H and R are checked against manual ruler measurements, and Q and tau are computed from independent timing—but it means the 'metric ruler' claim needs independent validation. Second, there are no error bars or significance tests anywhere. For a paper making claims as clean as these, that's a simple fix. Third, dataset and code are promised but not released; for a data-driven paper of this kind, that's a reproducibility gap. Fourth, the 'human-like capabilities' claim is not backed by a human experiment; psychoacoustic literature says humans can do this, but no comparison is made here. Minor: beta fixed at 0.62 is fine as an approximation, and the failure cases on hemispherical and multi-mode containers are disclosed, so that's not a hidden flaw.\n\nWho is this for? People working on audio-visual physical understanding, self-supervised learning, and robotic pouring. It's a paper I'd want to see after revision, not before. My recommendation: send it to serious reviewers, with the specific request that they verify the ground-truth source for l(t) and require independent validation of the dynamic metric. If that holds up, it's a strong paper.","headline":"A genuinely novel physics-grounded pipeline for pouring analysis, with a real but fixable circularity issue in the headline dynamic metric.","tokens_in":22799,"tokens_out":2952,"would_cite":true,"duration_ms":27113,"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":"From the sound of pouring alone, the paper recovers the air-column length, container height and radius, flow rate, and time to fill, by tracking the fundamental axial resonance whose wavelength is linear in the air-column length.","keywords":["pouring sounds","pitch detection","axial resonance","physical property inference","audio-visual co-supervision","container shape recognition","liquid mass estimation","fundamental frequency"],"falsifier":"Pour into a transparent cylinder with a ruler beside it at a known constant rate, record audio and video, extract the fundamental frequency from the spectrogram, and compare $l(t)$ from Eq. (5) against the visually tracked water level; if the discrepancy grows with $R$ or with flow rate, or if the apparent pitch is not single-valued, the fixed-$\\beta$ linear relation fails.","tokens_in":21670,"feed_emoji":"💧","tokens_out":7521,"duration_ms":66741,"temperature":0.7,"pith_summary":"The paper attempts to show that the sound of liquid pouring into a container carries enough information to recover the geometry of the container and the dynamics of the pour, without any visual input. Its route is a physical identity: the fundamental frequency of the air column above the rising liquid obeys $\\lambda(t)=4(l(t)+\\beta R)$ with $\\beta=0.62$, so the wavelength curve is a direct metric for the air-column length. From that curve, boundary conditions give container height and radius, and derivatives give flow rate and time to fill. The authors train a pitch detector without manual supervision, first on simulated pouring sounds and then on real recordings with a physics-inspired visual co-supervision, and report that the resulting model infers air-column length to 0.60 cm mean absolute error, height to 2.27 cm, radius to 1.39 cm, and flow rate to 22.5 ml/s on seen containers. A sympathetic reader would care because it turns an everyday, unlabeled sound into calibrated physical measurements and demonstrates generalization to unseen shapes, materials, and in-the-wild videos.","feed_headline":"Pouring sound is a metric ruler for the container","feed_subtitle":"One audio track gives air-column length, height, radius, and flow rate from the rising pitch.","key_machinery":"The load-bearing object is the axial-resonance wavelength identity $\\lambda(t)=4(l(t)+\\beta R)$, with $\\beta=0.62$ fixed; it converts pitch into a linear metric for the air column. Around this sit four derived formulas: $l(t)=(\\lambda(t)-\\lambda(T))/4$, $H=(\\lambda(0)-\\lambda(T))/4$, $R=\\lambda(T)/(4\\beta)$, $Q(t)=-(\\pi R^2/4)\\,d\\lambda/dt$, plus the early-pour approximation $\\tau(t)\\approx-\\lambda(t)/(d\\lambda/dt)$ for time to fill. The detector is an audio transformer trained in two stages: synthetic pre-training on simulated pours, then visual co-supervision on real pours, where a scale factor $\\alpha$ links metric wavelengths to pixel air-column lengths and radii. The wavelength curve carries the entire argument, because every physical property is a boundary value, an intercept, or a slope of that curve.","core_discovery":"The central claim is that the pitch of pouring water is the fundamental axial resonance of a pipe closed at one end, with a fixed end correction: $\\lambda(t)=4(l(t)+0.62R)$, where $\\lambda$ is the wavelength of the fundamental, $l(t)$ the air-column length, and $R$ the container radius. If this identity holds over the pour, then the wavelength curve is enough to read off $l(t)$ at every instant, $H$ and $R$ from the boundary at start and end, $Q(t)$ from its slope, and the time to fill from its early behavior. The paper supports the claim by building a transformer-based pitch detector that outputs a wavelength distribution per time step, pre-training it on synthetic pours generated with a differentiable synthesizer, and then fine-tuning it on real videos using the video stream as a weak teacher through the scale-aware equation $\\alpha\\lambda(t)/4 = l_{\\text{px}}(t)+\\beta R_{\\text{px}}$. Tested on a new dataset of 805 real pouring videos, the co-supervised model outperforms classical and learned pitch estimators and estimates physical properties with the errors reported above, while its features also support container-shape classification and liquid-mass regression on a previous dataset.","pith_inferences":["The paper leaves implicit that the same pitch-to-wavelength pipeline could serve as a contact-free calibration device: if the container is known, the formula gives flow rate and total poured volume, and if the flow rate is known, the same audio gives the container's dimensions; a direct test would compare audio-inferred volume against a scale.","A natural extension is multi-pitch tracking, since the failure on bottleneck containers with two simultaneous frequency modes suggests treating axial and radial resonances as separate tracks, or using a harmonic-aware architecture, rather than a single fundamental wavelength.","Because the relation is stated with a constant $\\beta=0.62$, the approach makes a testable prediction about how inferred size errors should scale with container radius; measuring that scaling would indicate whether the fixed end correction is an adequate approximation across the dataset's size range.","The result suggests that the metric-ruler claim may extend to other container-filling sounds, such as grains or viscous liquids, as long as an air column with a single dominant resonance exists; this is an extrapolation the paper does not make."],"forward_implications":["A single smartphone recording of a pour into a cylinder-like container yields absolute metric estimates of the container and the pour with no manual measurement, with reported mean absolute errors of 0.60 cm in air-column length, 2.27 cm in height, 1.39 cm in radius, and 22.5 ml/s in flow rate on Test set I.","Visual co-supervision improves estimates most near the end of the pour, where the audio signal is weak, and that is exactly where radius and flow-rate errors are dominated by $\\lambda(T)$ and the slope of $\\lambda$, so co-supervision translates into better static and dynamic properties.","Pitch-detection features encode shape: an unseen three-way shape classification reaches 90.91% sample accuracy and 92.47% mean class accuracy, and linear probing on a prior pouring dataset gives 1.20 oz mean absolute error for liquid mass.","The detector generalizes beyond cylinders to semi-conical, bottleneck, cup, teapot, and wine-glass containers, across glass, plastic, steel, ceramic, and cardboard, and to in-the-wild YouTube pours, although hemispherical containers and multi-modal bottleneck resonances are acknowledged failure cases."],"supporting_citations":[{"why":"Establishes the human perceptual ability to hear liquid level and time to fill, motivating the time-to-fill approximation.","marker":"[16]"},{"why":"Supplies the physics of vibrating shells and the radial resonance frequency used alongside axial resonance.","marker":"[33]"},{"why":"Provides the end-correction factor $\\beta=0.62$ that closes the wavelength identity $\\lambda(t)=4(l(t)+\\beta R)$.","marker":"[6]"},{"why":"Supplies the standing-wave analysis of a pipe closed at one end that the axial resonance formula rests on.","marker":"[79]"},{"why":"Provides the prior pouring dataset and supervised liquid-mass baseline against which the model is evaluated by linear probing.","marker":"[96]"},{"why":"Supplies the pretrained speech transformer architecture adapted for wavelength prediction.","marker":"[12]"},{"why":"Supplies pitch extraction for synthetic-data conditioning and is a comparison baseline.","marker":"[50]"},{"why":"Supplies the differentiable synthesis pipeline used to generate simulated pouring sounds.","marker":"[31]"},{"why":"Supplies the frozen self-supervised vision features used by the video co-supervisor.","marker":"[18]"},{"why":"Supplies container segmentation used to obtain pixel-level radius for scale-aware supervision.","marker":"[52]"}],"fun_headline_variants":["Pitch of pouring water reveals container size and flow","Sound of pour infers liquid level, shape, and rate","Water's pour tone is a built-in physics gauge","Hear the pour, read the cup: pitch predicts physics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that real pouring audio contains one clean, dominant axial-resonance pitch that follows $\\lambda(t)=4(l(t)+0.62R)$ at every instant, so every later measurement is a boundary value or slope of that curve; the paper itself shows hemispherical containers and some bottleneck containers violate this.","fun_headline_variants_meta":{"raw":{"variants":["Pitch of pouring water reveals container size and flow","Sound of pour infers liquid level, shape, and rate","Water's pour tone is a built-in physics gauge","Hear the pour, read the cup: pitch predicts physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1334,"prompt_tokens":979,"completion_tokens":355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":288}},"tokens_in":595,"tokens_out":355,"duration_ms":4219,"temperature":1.0,"reasoning_tokens":288,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:47:01.346985+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pour into a transparent cylinder with a ruler beside it at a known constant rate, record audio and video, extract the fundamental frequency from the spectrogram, and compare $l(t)$ from Eq. (5) against the visually tracked water level; if the discrepancy grows with $R$ or with flow rate, or if the apparent pitch is not single-valued, the fixed-$\\beta$ linear relation fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the physics of vibrating shells and the radial resonance frequency used alongside axial resonance."},{"cited_title":"The theory of sound, volume 2","cited_arxiv_id":null,"evidence_quote":"Supplies the standing-wave analysis of a pipe closed at one end that the axial resonance formula rests on."},{"cited_title":"Analyzing liquid pouring sequences via audio-visual neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the prior pouring dataset and supervised liquid-mass baseline against which the model is evaluated by linear probing."},{"cited_title":"Crepe: A convolutional representation for pitch estimation","cited_arxiv_id":null,"evidence_quote":"Supplies pitch extraction for synthetic-data conditioning and is a comparison baseline."},{"cited_title":"Ddsp: Differentiable digital signal processing","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable synthesis pipeline used to generate simulated pouring sounds."},{"cited_title":"Segment anything","cited_arxiv_id":null,"evidence_quote":"Supplies container segmentation used to obtain pixel-level radius for scale-aware supervision."}],"review_version":1}