{"id":"33268e19-1f3e-4a29-86b2-e35dfdfb1637","arxiv_id":"2501.17879","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The proposed NDPCA-aided distributed autoencoder with a task- and perception-aware loss achieves 19% and 52% PSNR improvements over naive distributed autoencoders for correlated speech, approaching a joint-encoder upper bound.","lead":"This paper proposes a compression method for speech recorded by several microphones and sent to a central receiver over a wireless link with limited bandwidth. It combines distributed PCA with a task- and perception-aware loss and reports large PSNR gains over naive autoencoder baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline PSNR gains are attributed to NDPCA exploiting inter-source correlation, but the correlation-exploitation mechanism is both underspecified and internally inconsistent: Eq.","rationale":"The reader's verdict is REJECT, and I agree that the central assertion is not adequately supported. The most load-bearing problem is not merely the absence of error bars or code, though those are real; it is that the mechanism claimed to produce the gains -- NDPCA exploiting cross-source correlation -- is never concretely realized in the paper. Section 3.2 gives only an SVD of each source's latent vector and says the system selects the 'B maximum correlation components' without specifying how components are ranked or how selection is trained. Section 4.1's cosine loss is internally contradictory: the prose says it penalizes high correlation, while Eq. (11) as written rewards it. Tables 2 and 3 then report loss values that, under the written metric, imply the source latents are nearly orthogonal, which would leave the distributed PCA with essentially no correlated structure to exploit. If that reading is correct, the 19%/52% improvements cannot be attributed to the proposed algorithm's central mechanism; if the implemented objective differs, the paper misdescribes its own method. The task-aware comparison is also confounded by the downstream SGMSE speech-enhancement step, but I focus on the correlation-objective issue because it undermines the foundational mechanism common to both the task-agnostic and task-aware claims. I mark agreement with the reader as partial rather than full because the reader's phrasing says Eq. (11) 'explicitly penalizes high correlation,' whereas the equation as written actually rewards it; the deeper issue is the unresolved inconsistency among the prose, the equation, and the reported table values. This does not change the verdict: the paper should not be accepted as evidence for the claimed NDPCA advantage until the loss and selection mechanism are corrected, implemented, and ablated.","tokens_in":12880,"tokens_out":7874,"duration_ms":81394,"concrete_test":"Re-implement the Section 3-4 pipeline on the CHiME-6 setup and log, every epoch, both the exact Eq. (11) value and the raw mean pairwise cosine similarity of the four source latent vectors, together with the number and identity of PCA components selected per source at inference. Then compare DPCA-E4D1 against naive E4D1 at the same total bandwidth. If the trained latents have mean pairwise cosine below about 0.1 and DPCA still beats naive E4D1 by roughly 19%, the gain is not caused by cross-source correlation; if the mean cosine is high (e.g., above 0.8), then the implemented loss cannot be Eq. (11) as written, and the paper's loss definition must be corrected before the claims can be evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claims (19% task-agnostic and 52% task-aware PSNR improvement, plus 91%/98.1% of the joint-encoder upper bound) are attributed to NDPCA selecting the 'B maximum correlation components' across source latents, as described in Section 3.2. For that mechanism to explain the gains, the trained source encoders must produce latent vectors that retain substantial cross-source correlation. That premise is not established and is contradicted by the paper's own reported losses. Section 4.1 states that the cosine term is intended to 'penalize high correlation,' but Eq. (11) defines Lcos = 1 - E[Σ_{i,j} z_i·z_j/(||z_i||||z_j||)], which is minimized by making latent vectors more correlated, not less. Tables 2 and 3 report 'Cosine Loss' values of 0.971 (E4D1, dim 8) and 0.767 (E4D1, dim 256); under the natural off-diagonal reading of Eq. (11), these values correspond to near-orthogonal source latents. Such latents contain almost no correlated principal components for the distributed PCA to select, so the claimed 18.8-19% advantage of DPCA over the naive distributed baseline cannot be explained by the proposed correlation-exploitation mechanism. If, instead, the implemented objective rewarded correlation, then the text is wrong and the reported loss numbers are not the training objective. Because the total loss combining Lcos with MSE, spectral SNR, task, and perceptual terms is never written, and no code or multi-seed results are provided, the objective actually minimized is unidentifiable from the paper. This is not a cosmetic typo: the central claim depends on a specific relationship between the training loss and the NDPCA selection, and the paper does not establish that relationship.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural distributed source coding scheme for correlated speech sources, combining distributed PCA (NDPCA) with task-aware and perception-aware losses to adaptively allocate bandwidth based on channel state information. The authors report 19% higher PSNR in task-agnostic settings and 52% higher PSNR in task-aware settings over naive distributed autoencoder baselines, and claim near-optimal performance relative to a joint-encoder 'upper bound.' The method uses variable-rate latent dimensions selected via distributed principal components and a loss comprising MSE, cosine similarity, spectral SNR, PSNR, a speech-enhancement task loss, and a perceptual loss.","tokens_in":13232,"tokens_out":6591,"duration_ms":57656,"significance":"If the reported results were reliable, the work would address a practically relevant gap: enabling dynamic bitrate adaptation for distributed speech coding with correlated sources while balancing task performance and perceptual realism. The paper is one of the first to apply learned distributed source coding to speech with a downstream enhancement task, and the use of CSI-derived bandwidth constraints is timely. However, the significance is substantially undercut by internal inconsistencies in the loss formulation, an unspecified core selection algorithm, a confounded task-aware comparison, and the absence of statistical validation. The claimed gains cannot currently be attributed to the proposed mechanism.","major_comments":[{"comment":"The cosine similarity loss is defined as Lcos = 1 - E[sum_{i,j} z_i . z_j / (||z_i|| ||z_j||)], which is minimized by increasing inter-source correlation, not by penalizing it as the text states. Tables 2 and 3 report 'Cosine Loss' values of 0.971 (dim 8) and 0.767 (dim 256) for E4D1; under the stated objective these imply near-orthogonal latents, leaving almost no correlated structure for the distributed PCA to exploit. This directly contradicts the mechanism claimed to produce the PSNR gains. The authors must clarify the actual implemented objective, define the total loss combining all terms, and state what the reported cosine values represent.","section":"Section 4.1, Eq. (11)"},{"comment":"The distributed PCA selection procedure is underspecified. The paper states that PCA (via SVD) is applied to each source's latent vector and that 'the B maximum correlation components from across all PCA components across all the sources' are selected, but it does not define the optimization problem, the measure of correlation between components, the selection algorithm, or how the decoder handles variable concatenation. Without this, the claimed 18.8% and 16.1% improvements over naive distributed baselines cannot be reproduced or attributed to NDPCA.","section":"Section 3.2"},{"comment":"The task-aware comparison is confounded by the downstream speech enhancement model. The pipeline applies a score-based SGMSE model to the decoder output before measuring PSNR, so the 52.2% improvement over the task-agnostic case may result from the enhancement model's denoising rather than from the proposed distributed coding or loss. A controlled ablation is needed, such as running the same enhancement model on the task-agnostic decoder outputs and comparing under identical conditions.","section":"Section 5.3, Fig. 3c"},{"comment":"The 'theoretical upper bound' (Joint E1D1) is not a theoretical bound; it is the authors' own baseline in which all source data are concatenated at a single encoder. No argument establishes that this architecture achieves the maximum possible PSNR for the given problem. Consequently, the statements that the proposed method reaches 91% (task-agnostic) and 98.1% (task-aware) of this 'upper bound' are self-referential and do not demonstrate near-optimality.","section":"Section 5.3"},{"comment":"Experimental reproducibility is inadequate for the headline claims. All results use a single fixed seed (seed=0), and no standard deviations, confidence intervals, or significance tests are reported for the PSNR gains of 18.8%, 16.1%, 52.2%, and 46.9%. Furthermore, the total loss combining MSE, cosine, spectral SNR, PSNR, task, and perceptual terms is never written down, so it is impossible to determine how the components were weighted or to re-run the experiments.","section":"Section 5.2 and 5.3"}],"minor_comments":[{"comment":"The worst-case Gaussian assumption for encoder outputs is introduced without justification; the rate-distortion bound in Eq. (1) applies to i.i.d. Gaussian scalar sources, and its application to arbitrary learned latents of dimension l_s,t is not defended.","section":"Section 3.1, Eq. (3)"},{"comment":"The speech enhancement loss is presented as the denoising score matching loss for the score network, but the role of this loss in training the encoders/decoders is not explained, especially whether gradients flow through the pre-trained score model and how the enhancement output is differentiated.","section":"Section 4.2, Eq. (17)"},{"comment":"The perceptual loss for the generator is not explicitly defined; Eq. (22) is the discriminator loss, and the text only mentions minimizing 'the difference of their logits' without giving the generator objective or the relative weight of the perceptual term.","section":"Section 4.3"},{"comment":"Table 1 appears empty in the submitted manuscript; it is referenced as 'Comparison of task-agnostic and task-aware loss' but contains no data.","section":"General"},{"comment":"There are several typos and inconsistencies: 'Spectogram' in Figure 1, 'to simply things' in Section 3.1, 'asign' in the Conclusion, and the abstract's '19%' versus Section 5.3's reported 18.8% and 16.1% gains (the abstract should clarify how the average is computed).","section":"General"},{"comment":"Reference 'han Li et al. 2023' should be formatted as 'P. Han Li et al.' or 'Li et al.' to match standard citation style; the duplicated Welker, Richter, and Gerkmann 2022a/2022b entries should be consolidated or distinguished.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript's central mechanism is internally inconsistent, and the main empirical claims are confounded and under-reported. These are not merely presentation issues: they require new experiments and likely a reformulation of the method. The self-referential 'theoretical upper bound' framing and the absence of any comparison with established learned distributed source coding baselines further weaken the contribution. If the authors can correct the loss definition, specify the NDPCA algorithm, remove the enhancement confound, and provide multi-seed statistics, a resubmission might be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on arXiv:2501.17879. The paper takes the NDPCA distributed source coding idea from Han Li et al. (2023) and applies it to multi-microphone speech, adding a speech enhancement task loss and an MS-STFT discriminator for perceptual quality. That application is new, and the dynamic bitrate adaptation by PCA-truncation is a practical approach for variable bandwidth. The experiments are framed correctly, with a joint encoder upper bound and naive distributed baselines, and the rate-distortion-perception trade-off curve is a nice addition.\n\nThe soft spots are real, though. The distributed PCA selection is never specified: how the 'B maximum correlation components' are chosen across sources, how each source's allocation is determined, and how the decoder gets the projections. None of that is described, and no code is provided. The results come from a single seed with no error bars, so the 19% and 52% PSNR gains are not statistically grounded. The task-aware comparison is also confounded: the task-aware PSNR is measured after a score-based enhancement model, not on the decoder output, so the 52% number mixes source coding gains with enhancement gains.\n\nThe most concerning issue is the cosine loss. The text says it penalizes high correlation between source latents, but Eq. (11) is 1 minus the expected cosine similarity, which rewards correlation. The reported cosine loss values in Tables 2-3 (0.971 for E4D1 at dim 8) don't fit either reading if self-pairs are included, which suggests the equation is incomplete or the implementation doesn't match the text. Since the NDPCA mechanism depends on exploiting inter-source correlation, the paper needs to clarify the actual objective and demonstrate that the trained latents retain correlation. As written, the mechanism is unidentifiable.\n\nI would still send this to reviewers. The system is plausible, the application is relevant, and the issues are fixable with a clearer loss formulation, more experimental details, and multi-seed runs. But the current claims should not be accepted as-is.\n\nRegards","headline":"Plausible NDPCA-for-speech extension, but the cosine loss inconsistency and missing experimental details undermine the central correlation-exploitation claim.","tokens_in":13783,"tokens_out":5910,"would_cite":false,"duration_ms":54010,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A34","94A15","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a neural distributed PCA-aided autoencoder with task- and perception-aware losses achieves up to 52% higher PSNR than naive distributed speech coding under bandwidth constraints, approaching the joint-encoder upper…","keywords":["distributed source coding","neural distributed PCA","speech enhancement","perception-aware loss","rate-distortion-perception trade-off","dynamic bandwidth","autoencoder","wireless AR/VR"],"falsifier":"Ablation: train the proposed model with and without the cosine-similarity loss, then measure the correlation among the PCA-selected components and the PSNR gap over the naive baseline; if removing the cosine term does not change the gap and the selected components are near-orthogonal, the claimed NDPCA mechanism is not what drives the results.","tokens_in":12704,"feed_emoji":"🎤","tokens_out":3918,"duration_ms":33220,"temperature":0.7,"pith_summary":"This paper tries to establish that distributed speech source coding can be made practically effective for bandwidth-constrained wireless links by combining neural distributed PCA with a task- and perception-aware loss. It claims that when several microphones record correlated speech, the proposed NDPCA-aided encoder-decoder outperforms naive per-source autoencoders by 19% in task-agnostic PSNR and by 52% when a downstream speech-enhancement task and perceptual realism are optimized, approaching the joint-encoder upper bound especially at low bandwidth. The paper also derives the uplink bitrate from channel state information so the model adapts to changing bandwidth without retraining. If true, this gives a concrete path for AR/VR telephony and multi-device speech pipelines where correlated recordings must be compressed in real time. A rate-distortion-perception trade-off curve is provided so operators can choose how much realism to sacrifice for distortion.","feed_headline":"Distributed speech coding beats naive codecs by 52% PSNR","feed_subtitle":"NDPCA-aided scheme approaches the all-sources-one-encoder upper bound at low bandwidth.","key_machinery":"Neural distributed PCA (NDPCA): after each source encoder produces a latent vector, PCA is applied per source and the B components with largest correlation across all sources are selected and transmitted, giving dynamic output dimensions that track channel capacity. The loss function couples this with a cosine-similarity term that penalizes inter-source correlation, a denoising score-matching loss from a score-based generative speech-enhancement model applied after decoding, and an MS-STFT discriminator perceptual loss that keeps reconstructed speech realistic.","core_discovery":"The central claim is that a distributed autoencoder whose per-source latent outputs are projected through neural distributed PCA, with a training loss that combines reconstruction error, a score-based speech-enhancement task loss, and a multi-scale STFT discriminator perceptual loss, achieves significantly higher reconstructed speech PSNR than naive per-source autoencoders under the same total bandwidth. In the task-agnostic setting the proposed E2D1 configuration reaches about 91% of the joint E1D1 upper bound, and in the task-aware setting about 98.1%, with the largest gains at low bandwidth. The paper presents these numbers as evidence that inter-source correlation can be exploited distributively without a shared encoder.","pith_inferences":["The cosine-similarity loss and the NDPCA component-selection step pull in opposite directions: the loss drives latent vectors toward orthogonality while the PCA step needs correlated components to exploit; ablating the cosine term and measuring the correlation of the selected components would test which side carries the result.","The 52% task-aware gain is reported against naive distributed baselines, not against a joint encoder with task loss; the meaningful comparison for practitioners is whether distributed gains persist when all baselines are trained with the same task and perception losses.","A direct extension would apply the same NDPCA-plus-task-loss scheme to other correlated modalities, such as multi-camera video or multi-sensor audio plus motion, where the downstream task defines the distortion metric."],"forward_implications":["Under bandwidth constraints, the NDPCA-aided distributed codec outperforms naive per-source autoencoders by 19% (task-agnostic) and 52% (task-aware) in average PSNR.","The proposed E2D1 variant approaches the joint E1D1 upper bound: 91% of its PSNR task-agnostically and 98.1% task-aware, especially at low bandwidth.","Channel-state-information-driven bitrate allocation lets the system vary encoder output dimensions at inference time without retraining for each output size.","The rate-distortion-perception trade-off curve shows that higher perceptual weights increase distortion at fixed bandwidth, giving operators a tunable knob for realism versus fidelity.","Task-aware training with the speech-enhancement loss substantially raises PSNR over task-agnostic reconstruction at the same bandwidth."],"supporting_citations":[{"why":"Supplies the NDPCA mechanism for distributed PCA component selection across correlated sources, which the proposed algorithm adapts.","marker":"(han Li et al. 2023)"},{"why":"Score-based Langevin diffusion model in the complex STFT domain used as the downstream speech-enhancement task.","marker":"(Welker, Richter, and Gerkmann 2022a)"},{"why":"MS-STFT discriminator used to compute the perceptual realism loss.","marker":"(D´efossez et al. 2022)"},{"why":"CHiME-6 dataset supplies the correlated multi-microphone conversational speech recordings.","marker":"(Barker et al. 2018)"},{"why":"Gaussian rate-distortion bound used to convert channel capacity into encoder output dimensions.","marker":"(Zamir, Kochman, and Erez 2008)"}],"fun_headline_variants":["NDPCA speech codec beats naive methods by 52% PSNR","Distributed speech coding nears single-encoder bound at low bandwidth","Perception-aware distributed speech codec gains 52% PSNR","Correlated speech codec: NDPCA approaches theoretical upper bound","Task-aware distributed speech coding: 52% PSNR boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that a training loss which penalizes correlation between source latents can coexist with a distributed-PCA step that needs correlated components to select; if the loss decorrelates the latents, the proposed advantage has no mechanism.","fun_headline_variants_meta":{"raw":{"variants":["NDPCA speech codec beats naive methods by 52% PSNR","Distributed speech coding nears single-encoder bound at low bandwidth","Perception-aware distributed speech codec gains 52% PSNR","Correlated speech codec: NDPCA approaches theoretical upper bound","Task-aware distributed speech coding: 52% PSNR boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":2988,"prompt_tokens":869,"completion_tokens":2119,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":2025}},"tokens_in":485,"tokens_out":2119,"duration_ms":14825,"temperature":1.0,"reasoning_tokens":2025,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:27:30.888125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablation: train the proposed model with and without the cosine-similarity loss, then measure the correlation among the PCA-selected components and the PSNR gap over the naive baseline; if removing the cosine term does not change the gap and the selected components are near-orthogonal, the claimed NDPCA mechanism is not what drives the results.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CHiME-6 dataset supplies the correlated multi-microphone conversational speech recordings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gaussian rate-distortion bound used to convert channel capacity into encoder output dimensions."}],"review_version":1}