{"id":"acd53777-9191-41fa-a663-add048b6148e","arxiv_id":"2412.07648","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A VAE trained on TF-IDF and Node2Vec embeddings of YAMNet audio events appears to cluster in-the-wild recordings by location, with only visual evidence from one user.","lead":"This paper presents a self-supervised pipeline that turns one-minute in-the-wild audio clips into latent vectors, and reports that a VAE's latent space separates indoor from metro recordings for one test user. A generalist might read it as an early step toward detecting risky or violent situations from ambient sound, but the evidence is qualitative and the dataset is private.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VAE separation claim rests on one user's qualitative heatmap and non-conclusive t-SNE; no quantitative or multi-user validation, so the central claim is unsupported as presented.","rationale":"The reader's weakest assumption correctly identifies GPS pseudo-label noise as a threat to the central claim, and the paper's own Section 6.1 supports that concern. However, the deeper and more load-bearing problem is that the central claim rests on qualitative, single-user visual evidence with no quantitative or statistical validation. Even if the pseudo-labels were perfect, the paper provides no measure showing that the VAE latent space separates acoustic scenes better than the raw embeddings or better than chance. The t-SNE plots are explicitly disclaimed as non-conclusive, and the heatmap 'clusters' are judged by eye. This makes the central claim unfalsifiable as presented. Consequently, the rejection verdict is appropriate; our concern reinforces it rather than altering it. We give credit for the transparent acknowledgment of label limitations and for the pipeline's potential as a starting point, but the evidence for the headline contribution does not meet the bar for acceptance.","tokens_in":12232,"tokens_out":2460,"duration_ms":22715,"concrete_test":"For every user, compute a quantitative clustering score on the held-out test latent vectors, e.g., the silhouette coefficient, using the GPS-cell/self-reported location tags as reference labels. Compare the VAE latent space against the raw TF-IDF+Node2Vec embeddings used as input. To account for label noise and the small number of clusters, generate a null distribution by permuting the location labels 1,000 times, preserving label frequencies, and compute a permutation p-value for each user. Report the median silhouette coefficient and median p-value across the 13 users. If the VAE's median silhouette is not significantly higher than the null distribution and not higher than the raw embeddings, then the claimed 'further discriminative latent spaces' are not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in Section 6.4, is that the VAE latent space 'is able to find a cluster related to metro locations and another cluster for indoor locations, leading to further discriminative latent spaces.' The evidence for this is visual: a cosine-distance heatmap for one user (Fig. 5) and t-SNE scatters for one user (Fig. 6). Section 5 explicitly notes t-SNE 'uses random initialisation, so these plots are not conclusive,' and Section 6.4 repeats that t-SNE initialization is random and leads to non-deterministic visual results. The heatmap clustering is interpreted by eye, with 'indoor' and 'metro' labels assigned to hexagons based on noisy GPS and self-reported labels, whose accuracy the authors themselves disclaim in Section 6.1 ('any outcome linked to an accuracy or any metric using these labels would not be significant'). Even if the labels were perfectly accurate, the claim that the VAE latent space is more discriminative than the raw embeddings is supported only by an observed 'larger range of similarity metric values,' which could be an artifact of the VAE's learned scaling rather than genuine separation of acoustic scene content. No quantitative clustering metric, no held-out multi-user evaluation, and no statistical test against chance structure is provided. The load-bearing assumption is therefore that a visually apparent cluster in one user's heatmap, colored by unreliable pseudo-labels, is trustworthy evidence of acoustic-scene discriminability; this assumption is untested and fragile.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a self-supervised pipeline for characterizing acoustic scenes in-the-wild from the WE-LIVE dataset. Audio is converted to discrete acoustic event activations using YAMNet, then to TF-IDF and Node2Vec embeddings, and finally a per-user linear VAE is trained to obtain a latent space. The authors claim that cosine-distance heatmaps and t-SNE visualizations of this latent space reveal distinct acoustic scenes, in particular separating indoor locations (home, work) from metro locations, whereas the raw embeddings do not. The main evidence is qualitative: one user's heatmap and t-SNE plots, with location pseudo-labels derived from GPS and self-reports.","tokens_in":12577,"tokens_out":3262,"duration_ms":31156,"significance":"If the central claim were properly supported, the paper would offer a useful exploratory result: a self-supervised VAE could impose structure on noisy, in-the-wild audio embeddings in a way that corresponds to meaningful acoustic environments, with potential applications to affective computing and risk detection. The work also contributes a description of the WE-LIVE dataset and a practical comparison of YAMNet and PANNs CNN14 inference times. However, the current evidence is qualitative and based on a single user, and the authors themselves state in Section 6.1 that the pseudo-labels are too unreliable for any metric-based evaluation. The significance is therefore conditional on substantial additional validation.","major_comments":[{"comment":"The central claim that the VAE latent space 'is able to find a cluster related to metro locations and another cluster for indoor locations' rests on visual inspection of one user's cosine-distance heatmap. Section 6.1 explicitly states that GPS-derived location tags are inaccurate and that 'any outcome linked to an accuracy or any metric using these labels would not be significant,' and Section 5 states that the t-SNE plots 'are not conclusive' due to random initialization. No clustering metric, statistical test, or multi-user replication is provided, so the evidence does not support the central claim as stated. The authors should either supply quantitative validation (e.g., silhouette score, adjusted mutual information, permutation tests against chance structure) on held-out data, or substantially temper the claim to a descriptive observation for an individual user.","section":"Section 6.4 and Section 5, Fig. 5"},{"comment":"The claim that the VAE latent space is 'more discriminative' than the raw embeddings is supported only by the observation of a 'larger range of similarity metric values.' This is not a valid measure of discriminability, because a VAE can rescale or rotate representations arbitrarily, and a larger range of cosine distances does not imply that the distances separate known acoustic scenes from one another. A quantitative comparison against the raw embeddings, using the same clustering or evaluation protocol on both representations, is needed.","section":"Section 5, Fig. 5"},{"comment":"The embedding pipeline depends on several user-specific and dataset-wide choices: the 99th percentile activation threshold for YAMNet probabilities, the hexagon grid edge of 0.0015, the restriction to the top 10 locations per user, the audio segment length of one minute, and the VAE hyperparameters. These are selected on the same data used for the qualitative evaluation, and no sensitivity analysis is reported. Because the activation threshold directly controls which acoustic events enter the TF-IDF/Node2Vec representation, the robustness of the claimed clustering to these choices should be demonstrated.","section":"Section 4.2 and Section 6.1"},{"comment":"The comparison between YAMNet and PANNs CNN14 is based solely on inference time, yet the paper later uses YAMNet outputs as the basis for the acoustic-scene representation. The choice of event detector could affect the resulting embeddings and the observed latent structure; without any evaluation of representation quality or downstream clustering behavior, the choice of YAMNet is not well justified for the paper's central goal.","section":"Section 6.2"}],"minor_comments":[{"comment":"Section 5 says the heatmap is shown 'for one user' and the t-SNE evolution 'for one user,' while Section 6.4 refers to 'the heatmap related to one user location (Fig. 5)' and 'another user' for Fig. 6; the text should clarify whether Figs. 5 and 6 are for the same user or different users.","section":"Section 5 and Section 6.4"},{"comment":"The VAE architecture is described only as 'linear' with ReLU activations and tanh on the last decoder layer; the number of layers, hidden dimensions, latent dimension, and early-stopping criterion are not specified, which limits reproducibility.","section":"Section 4.4"},{"comment":"There are several typos and incomplete references: 'recorede' in Section 3, 'Comparision' in Table 1, and references [4], [5], and [41] lack complete bibliographic information; the authors should also replace blog-style citations with primary sources where available.","section":"Throughout"},{"comment":"The sentence 'A 5x521 matrix is obtained related to Node2Vec embeddings' is unclear; the text should state explicitly that each of the 521 AudioSet events is represented by a 5-dimensional vector, and that the final embedding for a one-minute audio is a 6x521 matrix.","section":"Section 4.3"},{"comment":"The description of the WE-LIVE dataset mentions 14 female volunteers and that one was discarded due to insufficient audio data, but the number of users actually used in the experiments is not stated consistently; please state explicitly that 13 users were used in the analysis.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a preliminary exploratory study whose central conclusion is not yet supported by the presented evidence. The authors themselves acknowledge that the location pseudo-labels are too unreliable for metric-based evaluation, which makes the single-user qualitative heatmap insufficient. I would encourage the editor to seek a revised version with quantitative clustering validation, multi-user results, and sensitivity analysis; without such additions, the paper would not meet the bar for a serious journal. The citation list also contains several non-archival and blog-style sources that should be replaced with primary references."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an honest, clearly written exploratory paper whose central claim does not survive its own caveats. The authors build a sensible pipeline—YAMNet event probabilities binarized at a per-user 99th percentile threshold, TF-IDF plus ontology-derived Node2Vec embeddings concatenated, then a per-user linear VAE—and they run it on genuinely hard in-the-wild data, nearly a week of continuous wearable recordings from 14 women. The Node2Vec-on-AudioSet-ontology idea is the most interesting piece: it gives acoustic events semantic structure without hand-grouping the ontology. Credit where due: the paper reports that GPS labels are too unreliable for any accuracy metric, notes that t-SNE initialization is random and non-conclusive, and openly frames the work as a first characterization step. That candor is real.\n\nThe problem is the advertised result. The claim in Section 6.4 and the abstract, of 'distinct acoustic scenes' separating indoors from metro, rests on one user's cosine-distance heatmap and t-SNE scatters. No clustering metric, no statistical test against chance, no second user. Out of 13 usable users, that is a case study, not a finding. The authors' own Section 6.1 disclaims the very labels used to color the clusters, and a larger range of cosine values after the VAE could just be learned scaling rather than genuine separation. Hyperparameters (activation percentile, hexagon edge, latent dimension) are chosen on the same data with no sensitivity analysis. The abstract's 'pioneering' and 'stark contrast' outrun the hedged language in the body. The closely related work of Kushwaha et al. is cited but not compared in any detail. No code or data is released, so the YAMNet-versus-PANNs timing table is the only externally checkable number.\n\nWho is this for? Someone starting in in-the-wild acoustic scene characterization will find the pipeline write-up useful and the failure modes instructive. As a paper, I would not accept the central claim in this form. But I would send it to review rather than desk reject: the questions are legitimate, the writing is clear, the limitations are laid out, and the fixes are concrete—a silhouette or AMI score with a chance baseline, multi-user reporting, a permutation test, and a calibration check on the VAE's scaling.\n\nNet: engage with it if the venue tolerates heavy revision. It is a solid starting point, not a finished result.","headline":"Honest, clearly written exploratory pipeline, but the advertised VAE separation of indoor vs metro audio is one user's eyeballed heatmap colored by labels the authors themselves say cannot support any metric.","tokens_in":13091,"tokens_out":6292,"would_cite":false,"duration_ms":50605,"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":"The paper claims that a self-supervised variational autoencoder trained on sparse acoustic-event embeddings can regroup unlabeled in-the-wild audio by spatio-temporal acoustic scene, separating indoor locations from metro locations where…","keywords":["acoustic scene analysis","self-supervised learning","variational autoencoder","latent space","in-the-wild audio","acoustic event detection","TF-IDF","Node2Vec"],"falsifier":"Compute the cosine-distance matrix within and between location groups in the VAE latent space for all users and compare it to the same matrix computed after randomly shuffling the location labels; if shuffled labels produce clusters as separated as the real ones, the grouping is not driven by acoustic scene content. A second check: re-run the t-SNE visualization with several random seeds and see whether the indoor/metro split persists, since t-SNE initialization is nondeterministic.","tokens_in":12051,"feed_emoji":"🎧","tokens_out":5831,"duration_ms":48618,"temperature":0.7,"pith_summary":"This paper tries to establish that a variational autoencoder can turn unlabeled, in-the-wild audio into a latent space where acoustic scenes separate by their spatial context. Using one-minute recordings from a wearable dataset, the authors convert detected acoustic events into TF-IDF and Node2Vec embeddings, then train a per-user VAE on those sparse vectors. They report that in the latent space, the audio segments form clusters that line up with indoor locations (home and work) and metro locations, while the raw embeddings show no such grouping. If this holds, acoustic scenes can be characterized without clean location labels, which is a step toward detecting anomalous or violent situations from sound in real time.","feed_headline":"VAE separates subway audio scenes from indoor ones in the wild","feed_subtitle":"Everyday audio regrouped by acoustic context, a step toward detecting risk situations by sound.","key_machinery":"The load-bearing object is the per-user variational autoencoder and the embedding pipeline that feeds it. Audio is cut into one-minute segments, passed through YAMNet to get probabilities for 521 AudioSet event classes, binarized at a 99th-percentile threshold, and converted to a concatenated TF-IDF plus Node2Vec vector per segment. The VAE compresses these sparse vectors into a continuous latent space; cosine distance in that space and t-SNE visualizations are used to expose the cluster structure. Training uses stochastic gradient descent with momentum and an exponentially decaying learning rate to cope with the sparsity of the input matrices.","core_discovery":"The central discovery is that the VAE's latent space is more discriminative for acoustic scenes than the raw embeddings. For one user, the cosine-distance heatmap of latent vectors shows two clear clusters: one for indoor locations (home and work) and one for metro locations, whereas the heatmap of the input TF-IDF/Node2Vec embeddings is comparatively uniform. t-SNE projections over training epochs show the latent points gradually regrouping from a random scatter into location-associated clusters. The authors frame this as self-supervised spatio-temporal characterization: the model never sees location labels during training, and the noisy GPS-derived tags serve only as guides for interpreting the clusters.","pith_inferences":["A quantitative cluster-quality test (e.g., silhouette score computed on the latent vectors per user) would let the claimed indoor/metro separation be checked without relying on t-SNE's random initialization, which the paper itself flags as inconclusive.","If the effect generalizes across the 13 analyzed users, the same embeddings could serve as a pretrained representation for downstream audio-context classifiers, reducing the need for labeled risk-scene data.","The dependence of the clusters on YAMNet's 521 event classes could be probed by swapping in other AudioSet-based embeddings (e.g., PANNs or VGGish) and checking whether the same indoor/metro structure persists."],"forward_implications":["Acoustic scenes in the wild can be characterized without accurate location labels, since the VAE organizes the data before any label is used.","The latent space gives a more separable representation than the count-based and ontology-based embeddings alone, at least for the demonstrated user.","The same self-supervised pipeline could be applied to other users or datasets to obtain personalized scene models, and the per-user setup leaves room for personalization.","The authors position this characterization as the first step toward detecting anomalous or violent acoustic contexts in real time, with zero-shot extensions named as future work."],"supporting_citations":[{"why":"Supplies the YAMNet model that detects probabilities for 521 acoustic event classes in each audio segment.","marker":"[6]"},{"why":"The PANNs CNN14 model used as the comparison baseline for acoustic event detection inference time.","marker":"[7]"},{"why":"Defines the TF-IDF weighting used to turn acoustic event counts into one of the two embedding components.","marker":"[8]"},{"why":"Provides the Node2Vec algorithm that builds the semantic event embeddings from the AudioSet ontology.","marker":"[9]"},{"why":"AudioSet supplies the 521 event ontology and the pre-training data behind YAMNet's event classes.","marker":"[21]"},{"why":"The earlier affective acoustic scene study whose TF-IDF-based representations and cosine distance formulation this work extends.","marker":"[19]"},{"why":"Describes the WE-LIVE dataset: multimodal in-the-wild recordings with audio, GPS, and self-reported location and emotion labels.","marker":"[26]"},{"why":"Supplies the t-SNE algorithm used to visualize the latent space regrouping across training epochs.","marker":"[36]"}],"fun_headline_variants":["VAE latent space separates indoor from subway audio scenes","Self-supervised audio embeddings cluster by location in VAE","Acoustic scene separation emerges in VAE latent space","In-the-wild audio regrouping: VAE clusters indoor vs metro","VAE better at separating acoustic scenes than raw embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the noisy GPS-derived and self-reported location tags are accurate enough that the latent clusters can be interpreted as 'indoor' and 'metro' acoustic scenes; the paper itself says GPS is often inaccurate and that any metric using these labels would not be significant.","fun_headline_variants_meta":{"raw":{"variants":["VAE latent space separates indoor from subway audio scenes","Self-supervised audio embeddings cluster by location in VAE","Acoustic scene separation emerges in VAE latent space","In-the-wild audio regrouping: VAE clusters indoor vs metro","VAE better at separating acoustic scenes than raw embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1270,"prompt_tokens":951,"completion_tokens":319,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":239}},"tokens_in":567,"tokens_out":319,"duration_ms":3450,"temperature":1.0,"reasoning_tokens":239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:38:01.134806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the cosine-distance matrix within and between location groups in the VAE latent space for all users and compare it to the same matrix computed after randomly shuffling the location labels; if shuffled labels produce clusters as separated as the real ones, the grouping is not driven by acoustic scene content. A second check: re-run the t-SNE visualization with several random seeds and see whether the indoor/metro split persists, since t-SNE initialization is nondeterministic.","supporting_citations":[{"cited_title":"Plakal and D","cited_arxiv_id":null,"evidence_quote":"Supplies the YAMNet model that detects probabilities for 521 acoustic event classes in each audio segment."},{"cited_title":"Sparck Jones, ”A Statistical Interpretation of Term Specificity and Its Application in Retrieval,” Journal of Documentation, vol","cited_arxiv_id":null,"evidence_quote":"Defines the TF-IDF weighting used to turn acoustic event counts into one of the two embedding components."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AudioSet supplies the 521 event ontology and the pre-training data behind YAMNet's event classes."},{"cited_title":"Luis-Mingueza, E","cited_arxiv_id":null,"evidence_quote":"The earlier affective acoustic scene study whose TF-IDF-based representations and cosine distance formulation this work extends."},{"cited_title":"Rituerto Gonz ´alez, ”Multimodal Affective Computing in Wearable Devices with Applications in the Detection of Gender-based Violence,” Doctoral Dissertation, Dept","cited_arxiv_id":null,"evidence_quote":"Describes the WE-LIVE dataset: multimodal in-the-wild recordings with audio, GPS, and self-reported location and emotion labels."},{"cited_title":"van der Maaten y G","cited_arxiv_id":null,"evidence_quote":"Supplies the t-SNE algorithm used to visualize the latent space regrouping across training epochs."}],"review_version":1}