{"id":"1f170856-669a-4b9d-b73c-f8016a1cf4d4","arxiv_id":"2602.20967","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Mixing noisy and enhanced speech with weights derived from the recognizer's confidence on each signal reduces ASR word error rate without any additional training.","lead":"This paper shows that an ASR system's own confidence scores on the noisy version and the speech-enhanced version of an utterance can be used, at inference time and without extra training, to decide how to blend the two before recognition. The result is a simple drop-in post-processing step that lowers word error rate across several speech-enhancement-plus-ASR combinations and datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Confidence-ratio weights saturate near 0.5, so Conf-OA cannot down-weight a much worse enhanced signal; this explains the three Table-2 failures and weakens the 'reliable alternative' claim.","rationale":"The reader identified the confidence proxy as the weakest assumption, but focused on miscalibration in the direction sense (Table 3). That concern is largely mitigated by the OA structure: in miscalibrated cases, OA often still wins or ties. My concern is complementary but sharper: even when confidence ranks the two signals correctly, the magnitude of the ratio is compressed, so it under-weights the better signal in large-gap cases. The three Table-2 failures are direct evidence, and the WER-OA oracle shows these failures are not inherent to OA. This refines the reader's weakest assumption rather than replacing it. Since the reader's verdict is CONDITIONAL, and my concern adds a specific failure mode and a concrete test, the verdict remains CONDITIONAL—no change. I do not see a reason to move to REJECT: the method still improves over baselines in most configurations, and the failure cases are disclosed; however, the 'strong robustness' claim in the abstract is overstated. The proposed test could quantify the severity; if the deviation is small, the concern would be downgraded, but the absence of such a validation is exactly why a conditional verdict is appropriate.","tokens_in":9255,"tokens_out":7565,"duration_ms":71816,"concrete_test":"On CHiME-4 Real with Demucs and Whisper (the clearest failure case), compute for every utterance the confidence-based weight S'_conf from Eq. (3) and the WER-optimal weight S'_opt = WER(x̂)/(WER(y)+WER(x̂)) using ground-truth transcripts. Sort utterances by the WER gap |WER(y)-WER(x̂)| and plot the mean absolute deviation |S'_conf - S'_opt| in each decile. If the deviation grows with the gap—and S'_conf rarely leaves e.g. [0.3, 0.7] while S'_opt approaches 0 or 1—the saturation mechanism is confirmed. This directly tests the proxy's calibration magnitude without altering the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central premise (Sec. 2.1) is that ASR confidence, normalized via Eq. (3), yields an intelligibility-proportional weight. However, confidence scores are probability-like and saturate near the high end, so the ratio conf(y)/(conf(y)+conf(x̂)) is compressed toward 0.5 even when the WER gap is large. In contrast, the WER-based oracle Eq. (2) gives S' = WER(x̂)/(WER(y)+WER(x̂)), which can approach 0 or 1. Table 2 shows the consequences: for CHiME-4 with Demucs, the enhanced signal is dramatically worse than noisy (e.g., Real+Whisper: y=6.48, x̂=29.65). The oracle WER-OA achieves 5.97 in this same configuration, proving the OA structure can handle large gaps if the weight is set correctly. Conf-OA achieves only 6.60, worse than y, because its S' is too far from the optimal 0.82. The paper attributes these failures to 'the relative performance gap' (Sec. 4.1), but the oracle shows the gap is not the cause; the cause is the insufficiently extreme confidence ratio. Thus the load-bearing assumption that confidence is a reliable intelligibility proxy fails in precisely the high-gap regimes where the choice of weight matters most.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free Observation Addition (OA) method for noise-robust ASR. Given a noisy signal y and an SE-enhanced version xhat, it computes an utterance-level fusion weight S' = conf(y)/(conf(y)+conf(xhat)) from the frozen backend ASR's confidence scores (Eq. 3), then interpolates y and xhat (Eq. 1). Confidence is obtained from Whisper decoding log-probabilities or from Tsallis-entropy-based token posteriors for Parakeet/Wav2Vec2. Experiments cover two SE models (Demucs, GR-KAN MP-SENet), three ASRs, and two datasets, comparing against SNR-, DNSMOS-, and classifier-based OA baselines, plus a confidence-based hard switch and frame-level OA. The paper reports that Conf-OA generally outperforms existing practical OA baselines, approaches an oracle WER-based bound (Eq. 2), and honestly identifies three Table-2 configurations where it fails to beat the noisy input.","tokens_in":9492,"tokens_out":8967,"duration_ms":83190,"significance":"The contribution is practically interesting: if the result holds, it removes the trained predictor stage of prior neural OA methods and uses information already available from the ASR, easing deployment. The evaluation is broad and the oracle WER-OA provides a useful sanity check. The paper is honest about failure cases. However, the core premise that raw confidence ratios are a reliable intelligibility proxy is not independently validated; the saturation analysis in the major comments suggests the premise fails in large-gap regimes. The lack of statistical support further limits the strength of the empirical claims.","major_comments":[{"comment":"Eq. (3) defines S'=conf(y)/(conf(y)+conf(xhat)). With typical confidence scores both in [0.5,1], S' lies in [1/3,2/3], so it cannot weight a much worse enhanced signal near zero. In Table 2, the three cases where Conf-OA fails to beat Noisy y (CHiME-4 Simu+Whisper, Simu+Parakeet, Real+Whisper with Demucs) have large gaps; e.g., Real+Whisper y=6.48, xhat=29.65, yet WER-OA reaches 5.97 with S'=0.82. Thus the Sec. 4.1 explanation ('relative performance gap') is incomplete: the mechanism is confidence-ratio saturation. Please report confidence distributions, test a transform (e.g., logit) that can approach 0/1, and rerun these cases.","section":"Sec. 2.1, Eq. (3), Table 2"},{"comment":"Tables 1-4 give single WER values without error bars, seeds, or significance tests. Several key differences are tiny (Table 1 VoiceBank/Parakeet: Conf-OA 1.35 vs Classifier-OA2class 1.30; Table 2 VoiceBank/Parakeet: 2.11 vs 2.10). Without bootstrap confidence intervals or paired tests, 'best overall practical performance' (Sec. 4.1) may be noise. Please add statistical evaluation for Tables 1-2.","section":"Tables 1-4"},{"comment":"Sec. 2.1/2.3 introduce free parameters: Tsallis q=0.33, 'exponential normalization' temperature, CTC span pooling, epsilon. No values or sensitivity analysis are given. If these were tuned on the evaluation data, the 'training-free' claim is weakened; if fixed, state them. The asymmetry between Eq. (4) (token-weighted average for Whisper) and Eq. (5) (geometric mean for others) is also unexplained. This is needed for reproducibility.","section":"Sec. 2.1, Eq. (5); Sec. 2.3"}],"minor_comments":[{"comment":"Formatting errors in cells, e.g., '5.3724.85', '1.267.64', '2.762.11' are missing spaces between numbers.","section":"Tables 1, 2"},{"comment":"'V oiceBank' appears with an odd space; fix to 'VoiceBank'.","section":"Abstract / throughout"},{"comment":"The tie-breaking rule (choose y when confidences are equal) is arbitrary; since Table 3 shows a large ambiguous group, this choice could affect results and should be justified or tested.","section":"Sec. 2.2, Eq. (6)"},{"comment":"'TDT' is used without definition; also, the confidence computation requires access to internal decoding statistics (token log-probs, CTC alignments), which may not be available from black-box ASR APIs. This practical limitation is not discussed.","section":"Sec. 3.2.2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is substantiated: the confidence-ratio saturation appears to explain the three Table-2 failures better than the paper's stated reason. I do not recommend rejection, as the method is simple and broadly evaluated, but the authors should add a calibration/transform analysis and statistical evaluation before the central claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a short, honest paper with a simple idea: instead of computing the observation-addition weight from SNR, DNSMOS, or a trained classifier, take it from the backend ASR's own confidence (Eq. 3). I haven't seen that in the OA literature, and the empirical sweep is decent: two SE models, three ASRs, two datasets, five baselines. In most cells Conf-OA beats the noisy input and the enhanced input, and it beats the OA baselines. The training-free angle is real. I would send this to a referee.\n\nWhat is good beyond the headline: the WER-OA oracle (Eq. 2) is a nice reference point even though it needs ground truth; Table 3's miscalibration analysis shows OA helping precisely when confidence and WER disagree, which is a non-obvious benefit. The paper also discloses the three cases in Table 2 where Conf-OA does not beat noisy — that honesty matters.\n\nThe soft spots. The most important is the stress-test's point. The paper treats confidence as a practical approximation of intelligibility (Sec. 2.1). That approximation is miscalibrated in the high-gap regime. For CHiME-4 Real + Whisper with Demucs, noisy WER is 6.48 and enhanced 29.65. The WER-OA oracle gets 5.97 because it puts weight ~0.82 on noisy. Conf-OA gets 6.60 — worse than noisy. The paper says this is because the relative performance gap is large, but the oracle shows OA can handle that gap. The real issue is that confidence ratios do not go extreme enough: the confidence of a 29.65-WER transcript is not low enough relative to the 6.48 one. That is a load-bearing weakness for the paper's central claim that confidence is a reliable alternative. It works on average, but it fails in the very regime where the choice matters most.\n\nOther gaps are more conventional: no error bars, no significance tests, no code or data release, and the confidence hyperparameters (Tsallis q, temperature, epsilon) are hand-set without ablation. The frame-level comparison is brief and the conclusion is reasonable, but the frame-level result is hard to interpret without confidence visualizations.\n\nOverall, the contribution is real but modest. If a referee asks for a plot of S' vs optimal S' per utterance, or a calibration analysis, the paper would be substantially stronger. As-is, it is a borderline accept for a workshop or minor conference; for a journal it needs the missing evidential grounding. Send it to peer review — it deserves a serious referee, but I'd want the confidence calibration question addressed.","headline":"A simple, training-free OA weight from ASR confidence that mostly works; the confidence proxy is the soft spot, and the paper's own oracle shows it in the high-gap cases.","tokens_in":10117,"tokens_out":4349,"would_cite":false,"duration_ms":41888,"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":"This paper claims that a recognizer's own confidence scores can set the fusion weight for combining noisy and enhanced speech, yielding lower word error rates than existing observation-addition methods without any training.","keywords":["observation addition","speech enhancement post-processing","ASR confidence","noise-robust automatic speech recognition","training-free fusion","word error rate","speech intelligibility"],"falsifier":"Measure the Spearman correlation between the confidence difference conf(y) − conf(x̂) and the WER difference WER(x̂) − WER(y) over a large held-out noisy set; if the correlation is not positive, Eq. 3 systematically weights the worse signal. Concretely, if one constructs a set where the enhanced output is clearly better but the ASR is overconfident on the noisy input, Conf-OA should underperform simply choosing the enhanced signal, falsifying the claim that confidence is a reliable intelligibility proxy.","tokens_in":9040,"feed_emoji":"🎙️","tokens_out":2973,"duration_ms":33219,"temperature":0.7,"pith_summary":"Observation addition combines a noisy recording with a speech-enhanced version to reduce recognition errors, and its quality hinges on how the two are weighted. This paper proposes setting the weight by asking the backend ASR itself how confident it is in each signal, then forming the fused audio as a confidence-weighted sum. The claimed result is that this training-free rule (Conf-OA) lowers word error rate across two enhancement models, three recognizers, and two datasets, and beats prior OA methods that rely on SNR estimates, quality scores, or separately trained classifiers. The paper also shows an oracle version using true word error rates performs best, which supports the design while indicating confidence is only an approximation.","feed_headline":"ASR's own confidence can steer denoising fusion and cut word error","feed_subtitle":"No training or extra model: the recognizer's confidence sets the fusion weight and beats trained baselines across datasets.","key_machinery":"The key mechanism is confidence-weighted observation addition: a closed-form interpolation x̄ = S'·y + (1−S')·x̂ whose scalar S' is the ratio of the backend ASR's confidence in the noisy signal to the sum of confidences in the noisy and enhanced signals. Confidence is computed per ASR family—token-weighted geometric mean of segment log-probabilities for Whisper, and Tsallis-entropy-based token confidence for Parakeet and Wav2Vec2-CTC—so the same framework adapts to different recognizers without retraining. This identity converts an ASR's internal uncertainty into a fusion weight that favors the signal the recognizer is more likely to transcribe correctly.","core_discovery":"The central claim is that ASR confidence is a reliable, training-free substitute for true intelligibility in the observation-addition task. Using the fusion rule S' = conf(y) / (conf(y) + conf(x̂)), where y is the noisy signal, x̂ is the enhanced signal, and conf comes from the frozen backend ASR, the fused signal x̄ = S'·y + (1−S')·x̂ achieves the best practical WER across diverse SE-ASR combinations, outperforming SNR-based, DNSMOS-based, and trained-classifier OA baselines. The oracle WER-OA (Eq. 2), which uses true WERs, consistently reaches the lowest WER, confirming that intelligibility guidance is the right principle and that confidence is a workable proxy.","pith_inferences":["A natural extension is to generalize OA beyond two candidates—e.g., fusing one noisy signal with multiple enhanced versions—by normalizing confidence over all candidates, though the paper only tests the pairwise case.","Confidence calibration across ASR families is likely the main lever for closing the gap to oracle WER-OA; a lightweight calibration step (e.g., temperature scaling or Platt scaling on a small validation set) could improve Conf-OA without adding a full predictor.","Because the method depends on confidence being comparable for the same utterance's noisy and enhanced versions, it may be sensitive to systematic overconfidence in certain noise types; stress-testing on unseen noise distributions would reveal where the proxy breaks.","The frame-level negative result suggests that per-frame confidence estimates, while available, are not reliable enough to guide interpolation; sharper frame-level confidence (e.g., from a stronger CTC model) could be a testable path to fine-grained OA."],"forward_implications":["If Conf-OA holds up, speech enhancement front-ends can be post-processed without retraining or fine-tuning the SE or ASR models, making the method applicable to frozen, black-box systems.","ASR confidence can serve as a general proxy for intelligibility, meaning any ASR that emits confidence scores can drive observation addition without ground-truth transcriptions.","The method's wins over SNR- and DNSMOS-based OA suggest that targeting intelligibility (what the recognizer needs) rather than signal quality is the more effective fusion criterion.","Utterance-level fusion outperforms frame-level fusion, implying that preserving temporal continuity matters more than fine-grained adaptation in this setting.","The consistent superiority of oracle WER-OA indicates the headroom available to better calibrated confidence estimates."],"fun_headline_variants":["ASR's own confidence picks fusion weights, no training needed","Training-free fusion: ASR confidence beats trained baselines","Noisy ASR improved by confidence-guided observation addition","Intelligibility from ASR itself: training-free fusion wins","Let the ASR decide: confidence-driven denoising fusion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that ASR confidence for the noisy and enhanced signals is a faithful, same-scale estimate of each signal's expected recognition error—if the higher-WER signal receives higher confidence, the fusion weight points the wrong way.","fun_headline_variants_meta":{"raw":{"variants":["ASR's own confidence picks fusion weights, no training needed","Training-free fusion: ASR confidence beats trained baselines","Noisy ASR improved by confidence-guided observation addition","Intelligibility from ASR itself: training-free fusion wins","Let the ASR decide: confidence-driven denoising fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":953,"prompt_tokens":686,"completion_tokens":267,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":184}},"tokens_in":430,"tokens_out":267,"duration_ms":3821,"temperature":1.0,"reasoning_tokens":184,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:11:15.507642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the Spearman correlation between the confidence difference conf(y) − conf(x̂) and the WER difference WER(x̂) − WER(y) over a large held-out noisy set; if the correlation is not positive, Eq. 3 systematically weights the worse signal. Concretely, if one constructs a set where the enhanced output is clearly better but the ASR is overconfident on the noisy input, Conf-OA should underperform simply choosing the enhanced signal, falsifying the claim that confidence is a reliable intelligibility proxy.","supporting_citations":[],"review_version":1}