{"id":"46f4a7ae-a9ff-454f-86c1-a94f9856674c","arxiv_id":"2506.01138","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PARROT reports state-of-the-art speech emotion recognition by fusing Audio-MAMBA with attention-based SSL models using parallel Hadamard product and optimal transport branches.","lead":"PARROT fuses one Mamba-based and one attention-based speech model and reports top accuracy on three emotion recognition datasets. If the gains are real, it gives a cheap recipe for improving SER with frozen pre-trained models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's SOTA claim is contradicted by the paper's own Table 1 vs Table 2: the highlighted MESD fusion A(B)+M scores 69.05 while individual MMS scores 83.24 on the same protocol.","rationale":"The reader's weakest assumption was that the concatenation baseline is underspecified and that error bars/significance tests are missing. Those are valid concerns, but they are secondary. The more load-bearing problem is that the paper's central SOTA claim is contradicted by its own tables. On MESD, the individual attention-based PTM MMS achieves 83.24 accuracy in Table 1, while the highlighted heterogeneous PARROT fusion A(B)+M achieves only 69.05 in Table 2. Moreover, the best PARROT MESD entry is the homogeneous W2+M at 71.10, so the abstract's statement that PARROT beats individual PTMs and homogeneous fusion is false on that dataset. This is not a question of statistical significance or baseline tuning; it is a direct factual inconsistency. A reader cannot accept the headline claim without resolving this contradiction. If the numbers are typographical, the current manuscript is still unreliable until corrected; if they are not, the central claim is unsupported. For these reasons, the current version should not receive conditional acceptance; it should be rejected or sent back for major correction with verifiable per-fold results and code.","tokens_in":9630,"tokens_out":9591,"duration_ms":105489,"concrete_test":"Run the official PARROT code on MESD with Audio-MAMBA(base)+MMS under the stated five-fold protocol; verify whether the reproduced PARROT accuracy equals 69.05 and the reproduced individual MMS-CNN accuracy equals 83.24. Also compute the maximum PARROT MESD accuracy over all Table 2 combinations to check whether W2+M (71.10) is indeed the top entry and whether any PARROT fusion beats the individual MMS score of 83.24.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim in the abstract is that PARROT 'achieves SOTA results against individual PTMs, homogeneous PTMs fusion, and baseline fusion techniques.' The paper's own tables contradict this on MESD. In Table 1 (CNN downstream), the individual MMS model reaches 83.24 accuracy / 83.10 F1 on MESD, and Audio-MAMBA(base) reaches 78.03 / 77.96. In Table 2, the highlighted heterogeneous fusion A(B)+M via PARROT reaches only 69.05 / 68.72 on MESD, which is 14.19 accuracy points below the individual MMS model. The same Table 2 also shows that the highest PARROT MESD accuracy is W2+M at 71.10, a homogeneous attention-based fusion, not the selected A(B)+M entry. Thus the abstract's claim of superiority over individual PTMs and homogeneous fusion fails on MESD even using the paper's own reported numbers. This is not merely a missing baseline or missing error bars; it is an internal inconsistency between the headline claim and the experimental tables. Several other PARROT rows also underperform their best constituent (e.g., A(B)+U on all datasets), but the MESD case alone is sufficient to invalidate the SOTA claim as written. If the Table 1 or Table 2 MESD numbers contain a typo, that error itself undermines the reliability of all reported results until corrected and verified.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PARROT, a fusion framework for speech emotion recognition (SER) that combines one Mamba-based self-supervised pre-trained model (Audio-MAMBA) with one attention-based PTM. The framework applies parallel branches: a Hadamard product branch for element-wise local feature interaction and an optimal transport branch (Sinkhorn) for global distribution alignment, followed by concatenation and an FCN classifier. The authors evaluate on CREMA-D, Emo-DB, and MESD using five-fold cross-validation, compare against individual PTMs with SVM, FCN, and CNN downstreams, and against a concatenation baseline, and claim state-of-the-art results for heterogeneous fusion.","tokens_in":9914,"tokens_out":4963,"duration_ms":48344,"significance":"If the central claims held, the paper would be a useful empirical contribution on heterogeneous Mamba/attention fusion for SER, with a relatively lightweight fusion head (3.2M-13M parameters) over frozen PTMs and a public code repository. The methodology is not circular: the PTM features come from fixed external checkpoints, and the fusion head is trained on the target datasets. However, the reported evidence does not support the headline claim. The paper's own tables contain a direct counterexample on MESD, and the lack of statistical reliability measures makes it impossible to separate the reported gains from random variation.","major_comments":[{"comment":"The abstract claims SOTA results against individual PTMs, but on MESD the highlighted PARROT fusion A(B)+M reaches 69.05 accuracy, while the individual MMS CNN model in Table 1 reaches 83.24 on the same protocol, a 14.19-point deficit. Several other PARROT rows also underperform their best constituent (e.g., A(B)+U on MESD: 38.15 vs A(B) at 78.03; W2+M on MESD: 71.10 vs M at 83.24). This is not a missing baseline or missing error bar; it is an internal inconsistency between the headline SOTA claim and the reported numbers, and it directly contradicts the Section 4.2 statement that 'fusion of different PTMs through PARROT achieves better performance than the individual PTMs across all the datasets.'","section":"Section 4.2, Table 2 vs Table 1"},{"comment":"No confidence intervals, standard deviations, or significance tests are reported for the five-fold cross-validation. Many PARROT gains over the concatenation baseline are small (e.g., A(T)+H on CREMA-D: 66.82 vs 65.54; A(S)+W2 on CREMA-D: 62.34 vs 61.94), and PARROT is worse than concatenation on several entries (e.g., W2+M on MESD: 71.10 vs 72.14). Without error bars or paired tests, the claimed superiority of PARROT over baseline fusion techniques cannot be separated from random variation.","section":"Section 4.2, Table 2"},{"comment":"The concatenation baseline is described only as PARROT with the optimal transport and Hadamard product branches removed, with no details on how the remaining projections or classifier were tuned (e.g., learning rate, dropout, early stopping criteria), and it is not stated whether the same 120-dimensional projection is used. Since the central comparison is PARROT versus this baseline, the lack of an independently tuned standard fusion baseline weakens the empirical claim.","section":"Section 4.2, concatenation baseline"}],"minor_comments":[{"comment":"The text contains typos and awkward phrases: 'By the end of end of last decade' should be 'By the end of the last decade', 'conventational features' should be 'conventional features', and 'topmost most performance' should be 'topmost performance'.","section":"Introduction"},{"comment":"The sentence 'We representations are of dimensions...' should read 'The representations are of dimensions...'.","section":"Section 2"},{"comment":"The phrases 'This brings out limelight the effect' and 'In contrast, the some of the attention-based PTMs' are ungrammatical and should be revised.","section":"Section 4.2"},{"comment":"The Sinkhorn regularization parameter is not reported, although it is a free parameter that affects the transport plan and the results; please provide the value and any sensitivity analysis.","section":"Section 3.1"},{"comment":"The caption is dense and the labels U11, U22, U12, U21 are not defined in the main text; please define them explicitly in the caption or in Section 3.1.","section":"Figure 1"},{"comment":"It is not stated whether the five-fold cross-validation is speaker-independent; given that speaker overlap can inflate SER results, this should be clarified.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The central empirical claim is contradicted by the paper's own Table 1 and Table 2 on MESD. This is not a local issue that can be fixed by editing the abstract; the paper's stated contribution and the Section 4.2 conclusion would need to be substantially rewritten, and the reported MESD results actually undermine the hypothesis that heterogeneous fusion consistently improves over individual PTMs. Unless the authors can provide corrected experiments showing PARROT at least matches individual PTMs on all three datasets, with proper error bars and a well-specified baseline, the manuscript is not suitable for publication in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know one thing about this paper before reading further: the headline claim does not survive contact with its own tables. The abstract says PARROT achieves SOTA against individual PTMs and homogeneous fusion, but on MESD, Table 1 shows MMS alone at 83.24 accuracy with a CNN downstream, while Table 2 shows the highlighted PARROT fusion A(B)+M at 69.05. That is 14 points lower, and it is not even the best PARROT row on MESD—W2+M gets 71.10. The text explicitly calls A(B)+M the best MESD result, which is wrong on two counts.\n\nWhat is genuinely new: fusing Mamba-based and attention-based SSL representations for SER with a parallel Hadamard and optimal transport branch. I have not seen this exact combination, and the authors do cite the relevant prior work on homogeneous fusion and Mamba models. The setup is mostly sane: frozen external checkpoints, five-fold CV, three languages, and a concatenation baseline that removes the two branches while keeping the rest of the pipeline. On CREMA-D and Emo-DB, the A(B)+H fusion is plausibly complementary: 73.68 vs 70.63 concat and 92.24 vs 89.92. That is the kind of result that could matter for the SER subfield.\n\nBut the soft spots are not minor. The MESD contradiction is load-bearing because the abstract's SOTA claim is the paper's main contribution. The text also claims PARROT beats individual PTMs across all datasets, yet A(B)+U is below both constituents on every dataset (e.g., CREMA-D 62.53 vs A(B) 69.91). There are no confidence intervals or significance tests, so even the positive A(B)+H gain could be fold noise; many PARROT-vs-concat differences are around 1 point. The baseline description is thin on tuning details, and the Fp formula in Section 3.1 has an obvious typo (Rq→Rq instead of Rp→Rq). The code link in the PDF is mangled, so the promised reproducibility is not verifiable.\n\nOverall, the idea is worth a serious look, but the paper as written is not reliable. It needs a major revision: reconcile Table 1 and Table 2, report per-fold results with significance testing, tune and describe the baseline properly, and rewrite the claims to match what the data actually show. I would not cite it until those numbers are corrected. That said, it deserves a real referee rather than a desk reject—the fusion question is testable and the topic is relevant, and a careful referee could force the authors to fix the internal inconsistency rather than let it slide.","headline":"The fusion idea is new and the A(B)+H results are promising, but the paper's own Table 1 vs Table 2 contradicts its SOTA claim on MESD, and the missing error bars make the small gains unverifiable.","tokens_in":10488,"tokens_out":2832,"would_cite":false,"duration_ms":32704,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that PARROT, a parallel Hadamard-product and optimal-transport fusion of a Mamba-based and an attention-based self-supervised speech model, outperforms individual models, homogeneous attention-based fusion, and a…","keywords":["Speech Emotion Recognition","Self-Supervised Pre-Trained Models","Mamba State-Space Models","Attention-Based Models","Optimal Transport","Hadamard Product","Heterogeneous Fusion","Sinkhorn Algorithm"],"falsifier":"Run the same five-fold protocol with identical seeds and multiple restarts for PARROT and its concatenation-only ablation on CREMA-D, Emo-DB, and MESD, and report confidence intervals for accuracy and macro-F1; if concatenation matches or beats PARROT on any dataset, or if PARROT no longer beats the best individual PTM, the central claim fails.","tokens_in":9404,"feed_emoji":"🎭","tokens_out":9158,"duration_ms":78065,"temperature":0.7,"pith_summary":"The paper sets out to establish that Mamba-based and attention-based self-supervised speech models are complementary for speech emotion recognition, and that fusing the two families can beat fusing two attention-based models. It introduces PARROT, which combines a Hadamard-product branch for local element-wise interaction with an optimal-transport branch for global distribution alignment. Across CREMA-D, Emo-DB, and MESD, the best PARROT pairings improve on every individual pre-trained model and on a concatenation baseline, with top accuracies of 73.68, 92.24, and 69.05 respectively. If that holds, heterogeneous fusion is a practical way to push SER accuracy beyond what any single self-supervised model delivers.","feed_headline":"Fusing Mamba and attention models tops emotion-recognition tests","feed_subtitle":"A Hadamard and optimal-transport fusion block beats individual models on three speech emotion datasets.","key_machinery":"PARROT projects pooled last-hidden representations from two frozen PTMs through 1D convolutional blocks into a 120-dimensional latent space. The Hadamard branch computes $R_p \\odot R_q$ element-wise, preserving fine-grained local structure. The optimal-transport branch builds a cost matrix $C = \\|R_p - R_q\\|_2 / \\max(\\|R_p - R_q\\|_2)$, applies Sinkhorn to get a transport plan $\\Gamma = \\mathrm{Sinkhorn}(C)$, and transports each representation into the other's space; transported and original representations are then concatenated. The two branch outputs are concatenated and passed to an FCN classifier, so local interaction and global distribution alignment are combined in one network.","core_discovery":"On the paper's own terms, the core discovery is that fusing a Mamba-based SSL pre-trained model with an attention-based SSL pre-trained model through PARROT yields speech emotion recognition that outperforms every individual PTM tested, every homogeneous attention-based fusion, and every concatenation baseline. The strongest reported results are Audio-MAMBA(base) plus HuBERT at 73.68 accuracy / 72.90 macro-F1 on CREMA-D and 92.24 / 91.53 on Emo-DB, and Audio-MAMBA(base) plus MMS at 69.05 / 68.72 on MESD. The paper reads these results as evidence that the two architecture families capture complementary strengths, with attention models handling global dependencies and Mamba models handling efficient long-range processing, and that PARROT's two-branch design is what lets that complementarity emerge.","pith_inferences":["If the mechanism is general, the same two-branch recipe could be applied to other tasks where heterogeneous SSL PTM ensembles are used, such as speech recognition or audio deepfake detection; the paper only demonstrates it for SER.","The paper reports margins over concatenation of roughly 1 to 3 accuracy points with no variance or significance testing, so one natural next experiment is a multi-seed comparison with confidence intervals before treating the gains as stable.","An ablation the paper leaves implicit, running the Hadamard branch alone, the OT branch alone, and varying projection dimension or Sinkhorn regularization, would show which branch carries the gain; that is an editorial suggestion, not a paper claim."],"forward_implications":["SER systems can combine Mamba-based and attention-based SSL models instead of choosing one family, since PARROT improves over every individual PTM on all three datasets.","Reported reference points for future comparisons: Audio-MAMBA(base)+HuBERT gives 73.68 accuracy on CREMA-D and 92.24 on Emo-DB, while Audio-MAMBA(base)+MMS gives 69.05 on MESD.","Dataset-dependent pairing matters: MMS, the multilingual model, produces the best fusion only on the Spanish dataset, implying that PTM language coverage is part of the fusion decision.","Because PARROT beats homogeneous attention-based fusion, the paper's hypothesis that heterogeneous fusion is richer than homogeneous fusion is supported on these benchmarks."],"supporting_citations":[{"why":"Supplies Audio-MAMBA, the Mamba-based SSL PTM that forms one side of the heterogeneous fusion.","marker":"[13]"},{"why":"Defines the Mamba selective state-space architecture that Audio-MAMBA is built on.","marker":"[16]"},{"why":"Provides prior evidence that fusing SSL PTMs helps SER, the motivation for the fusion hypothesis.","marker":"[17]"},{"why":"Supplies HuBERT, the attention-based PTM in the top-scoring CREMA-D and Emo-DB pairs.","marker":"[23]"},{"why":"Supplies MMS, the multilingual attention-based PTM in the top-scoring MESD pair.","marker":"[24]"},{"why":"Provides the CREMA-D benchmark dataset used for the reported accuracy and F1 numbers.","marker":"[25]"},{"why":"Provides the Emo-DB German benchmark dataset used for the reported numbers.","marker":"[26]"},{"why":"Provides the MESD Mexican Spanish benchmark dataset used for the reported numbers.","marker":"[27]"}],"fun_headline_variants":["Fusing Mamba and attention SSL models boosts SER","Parallel fusion of Mamba and attention tops SER baselines","Mamba + attention fusion wins on three emotion sets","Hadamard transport fusion beats individual SSL models","PARROT fuses Mamba and attention for best SER results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the concatenation ablation is an honest control for PARROT; the paper gives no tuning details, error bars, or significance tests for that baseline, so the claimed superiority over baseline fusion could in principle be random variation.","fun_headline_variants_meta":{"raw":{"variants":["Fusing Mamba and attention SSL models boosts SER","Parallel fusion of Mamba and attention tops SER baselines","Mamba + attention fusion wins on three emotion sets","Hadamard transport fusion beats individual SSL models","PARROT fuses Mamba and attention for best SER results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1304,"prompt_tokens":890,"completion_tokens":414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":335}},"tokens_in":506,"tokens_out":414,"duration_ms":4140,"temperature":1.0,"reasoning_tokens":335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:49:05.491259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same five-fold protocol with identical seeds and multiple restarts for PARROT and its concatenation-only ablation on CREMA-D, Emo-DB, and MESD, and report confidence intervals for accuracy and macro-F1; if concatenation matches or beats PARROT on any dataset, or if PARROT no longer beats the best individual PTM, the central claim fails.","supporting_citations":[{"cited_title":"Ensemble deep learning with hubert for speech emo- tion recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies Audio-MAMBA, the Mamba-based SSL PTM that forms one side of the heterogeneous fusion."},{"cited_title":"Are paralinguistic representations all that is needed for speech emotion recognition?","cited_arxiv_id":null,"evidence_quote":"Defines the Mamba selective state-space architecture that Audio-MAMBA is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MESD Mexican Spanish benchmark dataset used for the reported numbers."}],"review_version":1}