{"id":"5cfc0750-33fa-4417-bae0-8f6ff6c8b65e","arxiv_id":"2501.02512","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-path Bi-Mamba network with temporal external attention achieves state-of-the-art RMSE and MAE for depression level estimation from long raw speech on AVEC2013/2014.","lead":"This preprint reports a neural network that estimates depression severity from long raw speech waveforms using a dual-path state space architecture with external attention. The authors report lower RMSE and MAE than prior systems on AVEC2013 and AVEC2014, with gains that grow as input length increases.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on borrowed baseline numbers that were not re-run under the paper's 30-second segment protocol; the reported RMSE gaps may be protocol artifacts rather than model gains.","rationale":"The paper's strongest claim is empirical: the proposed method achieves state-of-the-art depression-level estimation. For that claim to hold, the comparison numbers must be produced under comparable conditions. The manuscript itself flags that baseline results are adopted rather than re-run (Section III-D), and the evaluation protocol differs in several ways from typical prior work: raw waveform input, long segments up to 50 seconds, and segment-level RMSE where every segment inherits the full recording's BDI-II score. These differences are not minor; they change the effective test set and the difficulty of the task, so the reported margins over STA-EEP and TDCA-Net could be entirely due to protocol mismatch. This is a correctness risk, not a novelty or style issue. Tables II/III do provide same-pipeline comparisons against vanilla Transformer, Mamba, and DPRNN, and those comparisons are genuine evidence that the proposed dual-path Bi-Mamba plus temporal external attention helps for long-sequence regression on these datasets. However, those three baselines are generic sequence models, not prior depression-specific state-of-the-art systems, so they do not by themselves support the abstract's 'outstanding performance over the state-of-the-art.' The unexplained discrepancy between Table I and Table II/III for the proposed 30-second AVEC2013 result further weakens the empirical foundation, though it is not necessary to the main concern. Because the identified issue is addressable by re-running baselines and reporting variance, the reader's CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":6947,"tokens_out":6004,"duration_ms":61038,"concrete_test":"Re-run TDCA-Net and STA-EEP under the authors' exact protocol: 8 kHz raw waveform, 30-second segments, same train/dev/test folds, same per-segment BDI-II labels, and RMSE/MAE computed per segment. Use the original authors' code or checkpoints where available; otherwise provide a faithful reimplementation. Also report mean and standard deviation over at least 5 training runs for the proposed method and each baseline. If the matched-protocol baseline RMSE values remain above 9.24 on AVEC2013 and 9.20 on AVEC2014, the SOTA claim survives; if they drop to within noise or below, the headline comparison is a protocol artifact. In addition, rerun the proposed method at 30 seconds to determine whether Table I (9.24/8.26) or Table II (9.14/8.35) is the correct reported result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of state-of-the-art performance is not yet supported because the comparison against prior depression-estimation methods (Section III-D, Table I) uses externally reported numbers under uncontrolled protocols. The paper says 'we simply adopt the original results for two dataset baseline methods' and gives no evidence that STA-EEP [29] or TDCA-Net [14] were trained or evaluated with the same 30-second raw-waveform segments, 8 kHz resampling, train/dev/test split, or segment-level RMSE computed over duplicated BDI-II labels. AVEC2013/2014 recordings are 20-50 minutes long and are cut into 15/30/50-second segments, each inheriting the full-recording label; this changes the effective test set and sample weighting relative to prior per-recording or per-utterance evaluations. A method could appear better simply because the baseline numbers came from shorter segments, different features, or subject-level aggregation. The absence of error bars or significance tests compounds this problem. An internal inconsistency also exists: Table I reports the proposed 30-second AVEC2013 RMSE/MAE as 9.24/8.26, while Table II reports 9.14/8.35 for the same configuration, without explanation. The architecture is plausible, and Tables II/III provide same-pipeline comparisons against re-implemented Transformer, Mamba, and DPRNN baselines, but those do not establish 'state-of-the-art' over dedicated depression-estimation models under matched conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a time-domain, long-sequence speech model for estimating depression severity from raw waveforms. The architecture combines a dual-path structure in which the Bi-RNN of DPRNN is replaced by a bidirectional Mamba (Bi-Mamba), together with a temporal external attention module and a prediction module. Experiments on AVEC2013 and AVEC2014 use speech segments of 15s, 30s, and 50s, with each segment inheriting the full-recording BDI-II label. The paper reports RMSE/MAE improvements over Transformer, DPRNN, and Mamba baselines re-implemented under the same pipeline (Tables II and III), and claims state-of-the-art performance over prior depression-estimation systems (Table I), although the latter comparison relies on published baseline numbers adopted without re-running. The central claim is therefore an empirical benchmark comparison, with no formal derivation required for its correctness.","tokens_in":7254,"tokens_out":2237,"duration_ms":23476,"significance":"If the reported gains hold under a matched protocol, the paper would provide a useful data point for raw-waveform, long-context depression estimation, and the dual-path Bi-Mamba architecture is a plausible and timely combination of recent sequence-modelling ideas. The same-pipeline comparisons in Tables II and III against Transformer, DPRNN, and Mamba are a genuine strength, as is the explicit focus on long speech segments, which is relevant to clinical interaction settings. However, the headline state-of-the-art claim is not yet supported because the comparison in Section III-D/Table I uses externally reported baseline numbers under uncontrolled conditions, and the segment-level evaluation with duplicated full-recording labels complicates the interpretation of all RMSE/MAE numbers. The paper does not provide code, error bars, or significance tests, so the magnitude and reliability of the reported improvements cannot be assessed from the manuscript alone.","major_comments":[{"comment":"The state-of-the-art comparison is not a controlled comparison. The text states, 'Note that we simply adopt the original results for two dataset baseline methods,' meaning STA-EEP [29] and TDCA-Net [14] were not re-run under this paper's 30-second, 8 kHz, raw-waveform, segment-level duplicated-label protocol. Since the proposed method's advantage over these baselines (e.g., RMSE 9.24 vs. 10.54 for TDCA-Net on AVEC2013) may be wholly or partly an artifact of different segment lengths, features, or evaluation units, the 'outstanding performance over the state-of-the-art' claim in the Abstract and Section I is not established. The authors should either re-run prior methods under the exact proposed protocol or restrict the SOTA claim to the matched same-pipeline baselines in Tables II and III.","section":"Section III-D, Table I"},{"comment":"There is an internal inconsistency for the same proposed configuration at 30 seconds. Table I reports AVEC2013 RMSE/MAE of 9.24/8.26 for 'L + TEA', while Table II reports 9.14/8.35 for 'Proposed' at 30s; the AVEC2014 numbers also differ (9.20/8.39 vs. 9.20/8.39 in Table II? Actually Table II gives 9.20/8.39, matching, but the AVEC2013 row does not match). Since both tables should describe the identical model and protocol, this discrepancy needs to be explained, and the affected numbers corrected.","section":"Tables I and II"},{"comment":"Segment-level evaluation with duplicated full-recording BDI-II labels invalidates the assumption of independent test samples: multiple segments from the same recording are correlated and share the same label, so the effective sample size is much smaller than the number of segments. This affects all reported RMSE/MAE values and prevents direct comparison with prior per-recording or per-utterance protocols. The authors should either report results at the recording level (e.g., averaging segment predictions) or provide clustered error bars/significance tests (e.g., bootstrapping by subject or recording).","section":"Section III-C, evaluation protocol"},{"comment":"The paper does not report the model size, number of parameters, FLOPs, or runtime, despite the word 'Efficient' in the title and the claim of efficient long-sequence modelling. The choice of segment durations (15s, 30s, 50s) is also presented without a principled selection criterion or a held-out justification. Adding parameter counts and a complexity analysis would substantiate the efficiency claim, and a discussion of how segment lengths were chosen would strengthen the long-sequence argument.","section":"Section III-C and III-E"}],"minor_comments":[{"comment":"The spacing in 'A VEC2013' and 'A VEC2014' is inconsistent; the standard dataset names are 'AVEC2013' and 'AVEC2014'.","section":"Abstract and Section III-D"},{"comment":"The dimensions of the external attention maps are unclear. If Y ∈ R^{N×L}, then Mk and Mv are written as R^{N×L}, but then the operation Y × Mk in Eq. (8) is dimensionally ambiguous; please define the intended matrix multiplication and softmax axes explicitly.","section":"Section II-B, Eqs. (7) and (8)"},{"comment":"The description of AVEC2014 as 'derived from AVEC2013, includes the Northwind and FreeForm tasks' is vague; clarify how the 100-video merged set relates to the original 150 clips and to the standard train/dev/test splits.","section":"Section III-A"},{"comment":"Figure 3 is referenced in Section III-D but contains no error bars or confidence intervals, making the claimed 'consistent upward trend' difficult to evaluate; consider adding per-seed variability or at least stating the number of runs.","section":"Figure 3"},{"comment":"The table caption says 'AVEC 2014 DATASETS' but the table includes AVEC2013 columns; the caption should be corrected.","section":"Section III-D, Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper's central architectural idea is reasonable and the same-pipeline comparisons are informative, but the SOTA claim currently rests on non-matched baseline numbers and the evaluation protocol has a statistical flaw. I would not reject outright, because the issues are fixable with re-running baselines or softening the claim, adding recording-level evaluation, and reporting error bars. The internal Table I/II inconsistency must be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a reasonable engineering effort—DPRNN with Bi-Mamba plus temporal external attention on raw waveforms—and the same-pipeline comparisons against Transformer, DPRNN, and Mamba at 15/30/50s are worth seeing. The broader claim of state-of-the-art performance is not supported yet.\n\nWhat is actually new: applying a dual-path state-space architecture with external attention to depression level estimation from raw speech, and showing that the proposed model improves as segment length grows from 15s to 50s while the reimplemented baselines worsen or plateau. The ablation isolating external attention vs self-attention is useful. No new modeling principle emerges, but the application and the observation about long-sequence sensitivity are legitimate contributions.\n\nThe soft spots are in the evaluation. The comparison in Table I against STA-EEP and TDCA-Net merely copies published numbers without rerunning them, and the paper says so. Those numbers were likely produced under different segment lengths, features, and evaluation units, so the 'outstanding performance over SOTA' sentence is not supported. Also, the paper contains an internal inconsistency: Table I reports the proposed method's 30s AVEC2013 RMSE/MAE as 9.24/8.26, while Table II reports 9.14/8.35 for the same setting. That needs an explanation. There are no error bars or significance tests, and the segment-level evaluation duplicates each recording's BDI-II label across all its segments, which inflates the test set and correlates samples. This affects all methods, but it makes the reported gaps fragile. No code is provided, which limits reproducibility.\n\nThese are addressable issues. The core idea is plausible, and Tables II and III give useful same-pipeline baselines. I would send this to peer review with a strong request for revision: rerun the two borrowed baselines under the same protocol, add variance estimates, fix the table discrepancy, and discuss the label-duplication issue.\n\nFor whom: researchers in affective computing and digital health who want a time-domain long-sequence baseline. I wouldn't cite it in my own next paper, but I'd note it as a relevant recent result.","headline":"A useful engineering contribution with honest same-pipeline comparisons, but the SOTA claim rests on borrowed baselines and an unresolved table inconsistency.","tokens_in":7738,"tokens_out":2476,"would_cite":false,"duration_ms":25649,"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":"A dual-path bidirectional-Mamba network fed with up to 50 seconds of raw speech estimates depression severity more accurately than spectrogram-based and shorter-segment methods, achieving RMSE 9.05 on AVEC2013 and 9.14 on AVEC2014.","keywords":["depression","long sequence speech","time domain","state space model","Mamba","external attention","dual-path architecture","BDI-II"],"falsifier":"Re-run Transformer, DPRNN, and Mamba on AVEC2013 and AVEC2014 using exactly the same 8 kHz sampling, 15s/30s/50s segmentation, batch size 1, Adam with learning rate 0.002, and 100 epochs; if any baseline reaches or beats RMSE 9.05 on AVEC2013 or 9.14 on AVEC2014, the claimed superiority over the state of the art fails.","tokens_in":6768,"feed_emoji":"🎙️","tokens_out":7451,"duration_ms":62389,"temperature":0.7,"pith_summary":"Depression changes how people speak—slower, more paused, more monotone—and this paper argues that the way most automated systems listen throws those cues away. Standard approaches convert speech into spectrograms or Mel features, which discard information, and then chop recordings into short clips, which breaks the long-range patterns of a clinical interview. The paper proposes instead to feed up to 50 seconds of raw waveform directly into a network whose dual-path bidirectional-Mamba module reconstructs and models the long sequence, followed by a temporal external attention module that highlights depression-related cues. On AVEC2013 and AVEC2014, it reports RMSE 9.05 and 9.14, beating Transformer, DPRNN, Mamba, and published baselines, and, unlike those baselines, it keeps improving as the input gets longer.","feed_headline":"Longer raw speech sharpens depression severity estimates","feed_subtitle":"A dual-path Mamba network keeps improving as inputs stretch to 50 seconds, where Transformer, DPRNN, and Mamba stall.","key_machinery":"The load-bearing mechanism is the dual-path long-sequence modelling module built on Bi-Mamba. The raw waveform is reshaped into a 3D tensor of chunks; the intra-chunk block applies Bi-Mamba to each chunk, and the inter-chunk block applies it across chunks, and the two are fused and projected back to the original waveform shape. Bi-Mamba is a bidirectional variant of Mamba, a selective state space model whose parameters $\\Delta$, $\\hat{A}$, $\\hat{B}$, and $C$ are updated per input step, computed here with two parallel convolutions for forward and backward SSMs that are gated by the input and averaged. Around this core sits the temporal external attention module, which replaces the keys and values of self-attention with learned external matrices $M_k$ and $M_v$, capturing inter-instance correlations, and the prediction module that produces the final depression score.","core_discovery":"The central claim is that depression severity can be estimated from long raw speech waveforms in the time domain, without time-frequency transforms, and that the long-range structure of speech matters for accuracy. To show this, the paper constructs a network with three parts: a long-sequence modelling module that reshapes the waveform into chunks and applies bidirectional Mamba along both intra-chunk and inter-chunk axes, reconstructing the signal through a dual-path scheme borrowed from DPRNN; a temporal external attention module that uses learned external memory matrices in place of the keys and values of self-attention to emphasise depression-related cues; and a prediction module that maps the reconstructed sequence to a BDI-II score. The reported results are RMSE 9.05 on AVEC2013 and 9.14 on AVEC2014 with 50-second inputs, with the method outperforming vanilla Transformer, DPRNN, and Mamba at every tested length (15s, 30s, 50s) and outperforming published baselines in the 30s comparison.","pith_inferences":["If the long-context advantage holds, the same dual-path Bi-Mamba backbone could transfer to other paralinguistic assessments that suffer from short-window analysis, such as dementia or ADHD screening from speech.","The published baseline numbers were simply adopted rather than re-run under identical conditions; until those baselines are reproduced with the same 8kHz, 50-second segmentation and training recipe, part of the reported gap could reflect protocol differences.","A direct stress test would apply the model to an interview dataset such as DAIC-WOZ with 16kHz audio and PHQ-8 scores, checking whether the 8kHz, 50s design generalizes beyond AVEC.","The external attention module's learned memory size is a free parameter; varying it could reveal a bias-variance tradeoff in how much external context the depression estimator needs."],"forward_implications":["Longer input segments improve accuracy: the proposed model's RMSE on AVEC2013 drops from 9.50 at 15 seconds to 9.05 at 50 seconds, while the Transformer and DPRNN baselines stay flat or worsen.","Time-domain processing can replace time-frequency representations for depression estimation, avoiding information loss from Fourier and Mel transforms.","The gains come from both architectural components: the ablation shows that adding temporal external attention (L+TEA) outperforms adding self-attention (L+SA), and both beat the long-sequence module alone (L).","The method points toward a practical protocol in which a clinician's full interview segment, rather than a brief clip, is used for automated depression scoring."],"supporting_citations":[{"why":"supplies the dual-path RNN structure that the paper adapts by replacing Bi-RNN with Bi-Mamba for long-sequence modelling","marker":"[16]"},{"why":"provides Mamba, the selective state space model that Bi-Mamba is built from","marker":"[21]"},{"why":"provides external attention with learned keys and values, used in the temporal external attention module","marker":"[23]"},{"why":"provides the AVEC2013 dataset, its audio baseline, and the evaluation protocol","marker":"[24]"},{"why":"provides the AVEC2014 dataset and its audio baseline for evaluation","marker":"[25]"},{"why":"supports the claim that raw waveform processing can outperform time-frequency representations for depression prediction","marker":"[12]"},{"why":"a time-domain depression detection baseline whose published results are compared in Table I","marker":"[14]"},{"why":"the vanilla Transformer baseline compared in Tables II and III","marker":"[28]"}],"fun_headline_variants":["No Fourier: raw speech beats spectrograms for depression scores","Dual-path Mamba reads 50s speech to gauge depression","Long speech, no transforms: better depression estimate","Time-domain speech model outdoes spectrogram rivals","Depression cues from raw waveforms, no Fourier loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage over state-of-the-art methods relies on comparing against published baseline numbers that the authors did not re-run under the same audio lengths and training conditions; if those numbers came from a different setup, the improvement could be an artifact rather than real.","fun_headline_variants_meta":{"raw":{"variants":["No Fourier: raw speech beats spectrograms for depression scores","Dual-path Mamba reads 50s speech to gauge depression","Long speech, no transforms: better depression estimate","Time-domain speech model outdoes spectrogram rivals","Depression cues from raw waveforms, no Fourier loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1250,"prompt_tokens":951,"completion_tokens":299,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":220}},"tokens_in":567,"tokens_out":299,"duration_ms":3537,"temperature":1.0,"reasoning_tokens":220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:10:05.641515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Transformer, DPRNN, and Mamba on AVEC2013 and AVEC2014 using exactly the same 8 kHz sampling, 15s/30s/50s segmentation, batch size 1, Adam with learning rate 0.002, and 100 epochs; if any baseline reaches or beats RMSE 9.05 on AVEC2013 or 9.14 on AVEC2014, the claimed superiority over the state of the art fails.","supporting_citations":[{"cited_title":"Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech separation,","cited_arxiv_id":null,"evidence_quote":"supplies the dual-path RNN structure that the paper adapts by replacing Bi-RNN with Bi-Mamba for long-sequence modelling"},{"cited_title":"Beyond self-attention: External attention using two linear layers for visual tasks,","cited_arxiv_id":null,"evidence_quote":"provides external attention with learned keys and values, used in the temporal external attention module"},{"cited_title":"A VEC 2013: The Continu- ous Audio/Visual Emotion and Depression Recognition Challenge,","cited_arxiv_id":null,"evidence_quote":"provides the AVEC2013 dataset, its audio baseline, and the evaluation protocol"},{"cited_title":"A VEC 2014: 3D Dimensional Affect and De- pression Recognition Challenge,","cited_arxiv_id":null,"evidence_quote":"provides the AVEC2014 dataset and its audio baseline for evaluation"},{"cited_title":"Wavdepressionnet: Automatic depression level prediction via raw speech signals,","cited_arxiv_id":null,"evidence_quote":"supports the claim that raw waveform processing can outperform time-frequency representations for depression prediction"},{"cited_title":"Tdca-net: Time-domain channel attention network for depression detection.,","cited_arxiv_id":null,"evidence_quote":"a time-domain depression detection baseline whose published results are compared in Table I"}],"review_version":1}