{"id":"63948aa1-18b6-4730-b05c-a5aa9abe9fc8","arxiv_id":"2509.18010","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cross-attention in speech-to-text models correlates with saliency-based explanations (Pearson r roughly 0.49-0.75 in the best aggregations) but explains only a minority of the variance, so it should complement, not replace, attribution methods.","lead":"Speech-to-text models' cross-attention weights align only moderately with true input relevance, explaining well under half the signal when measured properly. The paper finds attention is an informative but incomplete explanation tool, cautioning against its use as a stand-alone explainability method in ASR and speech translation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline percentages rely on SPES as the sole reference for true relevance; an independent attribution baseline is needed to confirm the gap is real rather than an artifact of the reference saliency method.","rationale":"The reader's weakest assumption is that SPES saliency maps serve as a silver reference for true input and encoder-output relevance (Section 3.2, Appendix F). I agree that this is the most load-bearing assumption. The paper's deletion results provide some independent support for SPES at the input level, and the qualitative conclusion that cross-attention is incomplete is plausible and well supported by multiple analyses. However, the specific percentages in the abstract and Section 5.2—especially the 52–75% encoder-output claim—are not independently validated. If an independent attribution method produced much higher correlations with cross-attention, the central claim would need to be weakened or reframed. Because this is an unresolved empirical question, a conditional verdict is appropriate. I do not see a reason to move the verdict to accept or reject; the concern is addressable with additional experiments, not a demonstrated fatal flaw. The secondary issue of interpreting Pearson r as a percentage is real but does not change the direction of the qualitative conclusion, so it does not alter the recommended verdict.","tokens_in":26134,"tokens_out":8794,"duration_ms":86113,"concrete_test":"On a random subset of ~100 EuroParl-ST test segments from the base and small models, compute encoder-output saliency with an independent, non-perturbation attribution method (e.g., integrated gradients on encoder outputs H with a zero baseline, or expected gradients). Recompute the layer-average Pearson correlations between cross-attention and this independent SMH and compare them to Table 3's reported 0.52–0.75 range. If the independent correlations fall in the same range, the SPES-silver-reference concern is mitigated; if they differ by more than ~0.1 or exceed 0.8, the central quantitative claim is reference-dependent and should be re-estimated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that cross-attention captures only ~50% of input relevance and only 52–75% of encoder-output saliency—is measured entirely against SPES-based saliency maps (Section 3.2). The paper explicitly treats SPES as a 'silver reference' (Appendix F), but no independent attribution method is used to verify that SPES approximates true model relevance. This matters most for the encoder-output maps SMH: unlike the input-level SPES maps, which are partially validated by deletion scores (Appendix C, Table 5), SMH is an ad-hoc variant of SPES whose only validation is the dev-set deletion experiment in Appendix D. The headline '52–75% of saliency' claim in Section 5.2 therefore rests on the assumption that SMH faithfully reflects how the decoder uses encoder outputs. If SPES or its SMH variant is biased—e.g., over-weighting certain perturbation artifacts or missing relevance that cross-attention actually captures—the gap could be substantially smaller (or larger) than reported. A secondary issue is that the paper interprets Pearson r values directly as percentages of 'captured relevance' (e.g., r ≈ 0.75 becomes '75%'), which is not a valid variance-explained measure; correcting this would change the reported numbers, though it would likely strengthen rather than weaken the qualitative conclusion. The load-bearing concern is the reference-dependence of the measured gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether cross-attention scores in encoder-decoder speech-to-text (S2T) models can serve as explanations. It compares cross-attention to SPES-based input saliency maps and to an ad-hoc encoder-output saliency variant across monolingual ASR, multilingual/multitask ASR+ST models, at three scales and two languages. The main findings are that averaged cross-attention correlates moderately with input saliency (up to about 0.63) and more strongly with encoder-output saliency (up to about 0.75), but that cross-attention still 'captures only about 50% of input relevance' and only 52-75% of encoder-output saliency. The paper concludes that cross-attention is an informative but incomplete explanation proxy and should not be used as a stand-alone XAI tool.","tokens_in":26475,"tokens_out":3545,"duration_ms":31447,"significance":"If the headline quantitative claim is reliable, this is a useful and timely contribution: it directly tests a widely made assumption in speech-to-text research and has implications for timestamp estimation, alignment, and simultaneous translation. The experimental design has clear strengths: models are trained from scratch to avoid contamination; the analysis spans multiple tasks, languages, and scales; aggregation choices are ablated; and the deletion metric provides a behavioral check on explanation quality. The paper is also transparent about its limitations. The main concern is that the central percentages rest on a single attribution reference and on an invalid reading of Pearson correlation as a proportion of relevance.","major_comments":[{"comment":"The paper converts Pearson r values directly into percentages of 'total input relevance' and 'saliency' (e.g., '49-63%' and '52-75%' in Sections 5.1.2 and 5.2). A Pearson correlation is not a share of relevance; the proportion of shared variance is r^2. Under the reported numbers, r=0.63 corresponds to roughly 40% shared variance and r=0.75 to roughly 56%. Even r^2 would require the additional assumption that shared variance equals captured relevance. The quantitative headline in the abstract and conclusions is therefore not supported by the statistic used. Please reframe these percentages or replace them with a directly interpretable metric such as the deletion AUC already used in Appendix A. The qualitative conclusion may survive the correction, but the exact numbers must be revised.","section":"Section 5.1.2, Section 5.2, Tables 2-3"},{"comment":"The load-bearing claim that cross-attention 'captures only about 50% of the input relevance' is measured entirely against SPES, and the encoder-output claim is measured against SMH, an ad-hoc variant of SPES. SPES comes from the same research group (Fucci et al., 2025, with overlapping authorship), and SMH is validated only by the dev-set deletion experiment in Appendix D. If SPES or SMH is biased or incomplete, the measured gap between cross-attention and saliency is not a faithful measure of cross-attention's explanatory power. The limitation appendix acknowledges this in principle, but the abstract and conclusions present the gap as a property of cross-attention rather than as a comparison relative to one attribution method. Please add at least one independent attribution baseline (e.g., a gradient-based or perturbation method not developed by the same group) or a sensitivity analysis","section":"Section 3.2, Appendix F"},{"comment":"The 'context mixing' influence is quantified as '6.6-16.7%' from absolute differences in Pearson r between SMX and SMH. An absolute difference in correlation coefficients is not a percentage of explained behavior, and no derivation is provided to link a change in r to a fraction of context mixing. This should either be derived formally or, more conservatively, reported simply as a range of correlation improvements.","section":"Section 5.2, first paragraph"}],"minor_comments":[{"comment":"The title is typeset as 'CROSS-ATTENTION ISHALF EXPLANATION' in the preprint; there should be a space between 'IS' and 'HALF'.","section":"Title/Abstract"},{"comment":"The row for 'Whisper 10.69.0- 0.797 - -' appears malformed; the table columns do not align cleanly for the comparison models.","section":"Table 5"},{"comment":"Panel (i) is labeled 'CA(enASR)' but appears in the row for it-ASR; this is likely a typo and should be corrected.","section":"Figure 4"},{"comment":"There is a duplicated phrase: 'this pattern this pattern is absent' should read 'this pattern is absent'.","section":"Appendix E"},{"comment":"The captions read 'Person ρ' instead of 'Pearson ρ'.","section":"Tables 2 and 3"}],"recommendation":"major_revision","confidential_remarks":"The central measurement is heavily dependent on SPES, which is from the same group; this is a circularity risk that should be surfaced in the review. The paper is otherwise methodologically careful, and the qualitative conclusion is likely robust, but the quantitative headline needs either an independent reference method or a more cautious framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Brief take: this paper gives the first systematic empirical answer to whether cross-attention explains predictions in speech-to-text models, and the answer is \"partially, but not enough to trust it alone.\" It's worth reading and refereeing. The main caveats are that the headline numbers treat Pearson correlation as if it were explained variance, and the reference saliency maps come from the same research group, so the measured gap is against a possibly biased yardstick.\n\nWhat's genuinely new: nobody has done this breadth of comparison in S2T before. They train models from scratch to avoid contamination, sweep monolingual/multilingual, ASR/ST, two model sizes, and back the correlation results with a deletion metric. The consistent finding—that averaged cross-attention correlates moderately with saliency but falls well short of full relevance—is credible. The aggregation ablations and the context-mixing angle (input vs. encoder-output saliency) are thoughtful additions. They also openly acknowledge the silver-reference assumption in the limitations section, which is more than many papers do.\n\nThe soft spots are real but fixable. First, Sections 5.1.2 and 5.2 interpret Pearson r values of roughly 0.5–0.75 as capturing \"49–63%\" or \"52–75%\" of relevance. That's not what r measures; r² would give 25–56%. The qualitative conclusion doesn't change—the correlation is clearly incomplete—but the numbers as percentages are misleading. Second, the reference explanations come from SPES, which shares authors with this paper, and the encoder-output variant SMH is an ad-hoc extension validated only on a dev-set deletion experiment. So the exact size of the gap is method-dependent. An independent attribution method would firm up the claim. The deletion analysis helps here: cross-attention's deletion score (41.2) is much worse than full saliency (91.3), which supports the \"informative but incomplete\" framing without relying on correlation-to-percentage conversion.\n\nThis paper is for anyone using cross-attention for timestamping, forced alignment, or simultaneous translation in speech. It should reset expectations and push toward either better aggregation or attention regularized toward saliency. I'd send it to peer review: the core message is likely to hold, and the numeric overreach and reference-dependence are addressable in revision.","headline":"First systematic test of cross-attention as explanation in speech-to-text; the qualitative conclusion is solid, but the headline percentages overstate the precision and the reference saliency method is not independent.","tokens_in":26960,"tokens_out":1583,"would_cite":true,"duration_ms":15391,"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":"Cross-attention scores in encoder-decoder speech-to-text models capture only about half of the input relevance identified by perturbation-based saliency maps, so the paper argues they cannot stand alone as explanations.","keywords":["cross-attention","speech-to-text","explainable AI","saliency maps","feature attribution","context mixing","automatic speech recognition","speech translation"],"falsifier":"Run the same correlation and deletion pipeline with an independent reference—manually aligned word or phoneme boundaries, or a second feature-attribution method—and check whether cross-attention matches that reference far better than it matches the paper's saliency baseline. In particular, an attention-based deletion score near the full-resolution saliency score (91.3 on the paper's base model, versus 41.2 for attention) would falsify the 'half explanation' conclusion.","tokens_in":26040,"feed_emoji":"🎙️","tokens_out":8924,"duration_ms":80025,"temperature":0.7,"pith_summary":"The paper tests a routine assumption in speech-to-text models: that the cross-attention weights connecting the decoder to the encoder show which parts of the audio actually drove each generated word. It compares those weights against saliency maps produced by a perturbation-based feature-attribution method, across monolingual and multilingual models, ASR and speech-translation tasks, and three model sizes. It finds that cross-attention correlates moderately to strongly with input saliency, especially when scores are averaged over heads and layers, but still accounts for only about half of the input relevance. Even when compared to saliency computed directly on encoder outputs, the numbers reach only 52–75% of what the saliency reference shows. The conclusion is practical: cross-attention is a cheap, partial signal that can complement attribution methods, but it should not be treated as a stand-alone explanation, especially in applications like timestamp estimation and alignment.","feed_headline":"Cross-attention explains only half of speech-to-text decisions","feed_subtitle":"Attention scores are only partial evidence of what audio drove each word; tools built on them need a stronger signal.","key_machinery":"The load-bearing comparison objects are the cross-attention matrix, extracted from each decoder layer and head during generation, and two saliency references: input saliency maps, obtained by perturbing spectrogram regions and measuring KL divergence in the output distribution, and encoder-output saliency maps, computed by perturbing encoder hidden states. The cross-attention matrix is averaged over selected heads and layers; saliency maps are aggregated across frequency and time to the same temporal resolution. The correlations between flattened attention and saliency vectors are measured with Pearson correlation, and the difference between input-level and encoder-output-level correlations","core_discovery":"Central claim: in encoder-decoder speech-to-text models, cross-attention is only a partial explanation. On the monolingual ASR model the best head/layer aggregation correlates at about 0.59 with input saliency; on multitask multilingual models layer-averaged correlations run 0.49–0.63, which the paper reads as 49–63% of input relevance. Against encoder-output saliency—where context mixing is no longer a confound—correlations rise by 0.03–0.18 but plateau at 52–75%. A deletion check on the base model sharpens the point: cross-attention scores 41.2, frequency-aggregated saliency 52.9, full-resolution saliency 91.3. The paper takes this as evidence that cross-attention is an auxiliary cue, not","pith_inferences":["A plausible editorial extension: if the ~50% gap holds, attention-map visualizations shipped in real products (e.g., timestamped transcriptions) should carry a confidence caveat and be cross-checked against forced alignment rather than presented as ground truth.","A natural extension the paper mentions but does not test is to train with an auxiliary loss that pulls cross-attention toward saliency maps or monotonic alignment; the results here predict that such regularization would improve both faithfulness and downstream word-level accuracy.","The paper deliberately excludes speech-LLM architectures; since those models add a second encoder-decoder boundary, the context-mixing effect measured here could be larger there, making cross-attention an even weaker explanation for that family."],"forward_implications":["Downstream tools that use cross-attention for timestamp prediction, audio-text alignment, or guided simultaneous translation should stop relying on a single head or layer; averaging over heads and, where possible, layers is closer to the saliency reference.","Because cross-attention aligns better with encoder-output saliency than with input saliency, attention-based explanations are better read as descriptions of how the decoder uses encoded representations, not of which raw audio regions matter.","Cross-attention cannot replace feature-attribution methods for faithful explanations; its upper range of 52–75% against encoder-output saliency leaves too much unexplained for it to stand alone.","The last decoder layers are consistently the most aligned with saliency across tasks, languages, and model sizes, so layer choice is consequential if attention is used at all."],"fun_headline_variants":["Cross-attention: only half the explanation for speech-to-text","Speech-to-text models: attention scores explain only half","Cross-attention covers half of speech-to-text input relevance","Attention scores capture just half of speech-to-text reasoning","In speech-to-text, cross-attention is a partial explanation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central numbers depend on treating the perturbation-based saliency maps as a silver reference for what the model truly uses; if that reference is biased, the measured gap could be an artifact of the reference rather than a property of cross-attention, as the paper itself acknowledges in its limitations section.","fun_headline_variants_meta":{"raw":{"variants":["Cross-attention: only half the explanation for speech-to-text","Speech-to-text models: attention scores explain only half","Cross-attention covers half of speech-to-text input relevance","Attention scores capture just half of speech-to-text reasoning","In speech-to-text, cross-attention is a partial explanation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00072,"raw_usage":{"total_tokens":3093,"prompt_tokens":789,"completion_tokens":2304,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2233}},"tokens_in":533,"tokens_out":2304,"duration_ms":16083,"temperature":1.0,"reasoning_tokens":2233,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:45:58.856757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same correlation and deletion pipeline with an independent reference—manually aligned word or phoneme boundaries, or a second feature-attribution method—and check whether cross-attention matches that reference far better than it matches the paper's saliency baseline. In particular, an attention-based deletion score near the full-resolution saliency score (91.3 on the paper's base model, versus 41.2 for attention) would falsify the 'half explanation' conclusion.","supporting_citations":[],"review_version":1}