{"id":"95525562-63ad-407e-a766-c7813b05256d","arxiv_id":"2607.03670","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Multi-model ASR consensus (BEACON) curates 413 h of CHILDES with corrected timestamps; the 283 h ASR subset yields up to 19.5% relative WER reduction on four held-out child benchmarks.","lead":"BEACON recovers utterance timestamps in long child-speech recordings by aligning multiple off-the-shelf ASR word streams to trusted transcripts and taking consensus. The resulting 283-hour CHILDES subset cuts out-of-domain child ASR error by up to 19.5% relative.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged proxy limitation; the strongest claim is the downstream WER result, which stands independently.","rationale":"The reader's CONDITIONAL verdict already isolates the correct soft spot (proxy without absolute boundary ground truth) and correctly treats the multi-model out-of-domain WER reductions as real and checkable given the public release. My re-read finds no stronger load-bearing flaw that would move the verdict toward REJECT or require a different condition. The downstream experiment uses a separate quality filter (verbatim normalization + independent ASR agreement) and three model families, so even if the Granite proxy shares bias with the ensemble, the training utility claim remains supported by Table III. The concrete test above would cleanly separate 'timestamps are accurate' from 'clips are useful for ASR,' which is exactly the distinction the paper already draws between Sections V-A and V-B. No change to CONDITIONAL is warranted.","tokens_in":15618,"tokens_out":539,"duration_ms":5352,"concrete_test":"On a stratified 50-recording subsample that already has human-corrected utterance times (or that can be re-annotated), compute mean absolute onset and offset error of BEACON vs. raw CHILDES vs. BatchAlign2; if BEACON's absolute error is not lower while Table III gains remain, the proxy concern is confirmed but the strongest claim is unaffected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is the multi-model fine-tuning result in Table III (average relative WER reductions of 19.5%/3.6%/5.6% on four held-out child benchmarks). That claim does not rest on the Section V-A ASR-proxy being a perfect measure of absolute boundary error; it rests on the released 282.6 h clips being usable supervised pairs. The reader's weakest_assumption correctly notes that the proxy never reports onset/offset error against human times and could share systematic bias with the ensemble. That is a real limitation for any claim that BEACON recovers 'accurate' timestamps, but it is not load-bearing for the headline downstream result: the same clips that pass the independent WER-based misalignment filter (ins/del ≤ 0.25) demonstrably improve three distinct ASR families on non-overlapping data. Hyper-parameters set by light inspection and abstention-induced selection bias are acknowledged in Section VI and likewise do not overturn the reported Δ WER numbers. No internal inconsistency or hidden assumption that would reverse Table III was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces BEACON, a corpus-agnostic pipeline that recovers utterance-level timestamps for long-form audio paired with a trusted transcript by (i) decoding with four diverse off-the-shelf ASR systems, (ii) aligning each model’s word stream to the reference via chunked edit-distance search and monotonic DP, and (iii) fusing per-model spans by majority consensus (VERIFIED / RECOVERED / UNRESOLVED). Applied to English CHILDES, the authors release a 413-hour general-purpose set that preserves raw CHAT annotations and a stricter 283-hour ASR-training subset with verbatim normalization and insertion/deletion filtering. Timestamp quality is assessed via a bounded-WER proxy on clips cut by each method (Table II); downstream usefulness is shown by fine-tuning Whisper, Parakeet, and Canary on the ASR subset and evaluating on four non-overlapping child benchmarks (RSR, MyST, OGI Kids, CMU Kids), reporting average relative WER reductions of 19.5%, 3.6%, and 5.6% respectively versus zero-shot (Table III).","tokens_in":15914,"tokens_out":1195,"duration_ms":17451,"significance":"Child ASR remains data-scarce; a large, publicly released, speaker-attributed CHILDES-derived resource with corrected timestamps is of clear practical value. The multi-model consensus design is a sensible, reusable recipe for any long-form corpus with trusted text but unreliable times, and the authors ship code and curated data. The downstream experiment is well controlled: three model families, a high-precision low-yield FASA baseline, and four held-out benchmarks that do not overlap training. If the released clips remain usable supervised pairs—as Table III indicates—the work supplies a concrete training resource and a general timestamp-curation method that other labs can apply. Strengths to credit explicitly: public data/code release, multi-family OOD evaluation, and an independent proxy ASR (Granite) outside the ensemble.","major_comments":[{"comment":"Section V-A and Table II: the claim that recovered timestamps are “more accurate” rests solely on a bounded-WER proxy from a fixed external ASR. Lower proxy WER is consistent with better audio–text agreement but does not measure absolute onset/offset error against human-annotated times; any systematic bias shared by the proxy and the ensemble could inflate apparent quality. This is load-bearing for the accuracy narrative of BEACON itself (as distinct from the downstream usefulness claim in V-B). The manuscript should either (a) report a small human re-annotation sample with absolute boundary error, or (b) substantially strengthen the caveats so that Table II is framed only as a relative ranking of clip usability, not as absolute timestamp accuracy.","section":"Section V-A, Table II"},{"comment":"Section III-B/C and Section VI: free parameters (L_chunk, τ_wer, ε/θ/κ, P_miss, p_tok, w_jmp, insertion/deletion threshold 0.25, etc.) are set by light manual inspection with no ablation or sensitivity analysis. The Limitations section acknowledges this, but the operating point is still presented as the default recipe. Because yield, RECOVERED/UNRESOLVED rates, and the ASR-subset size depend on these choices, a minimal sensitivity study (or at least reporting status-label counts and yield under a few nearby settings) is needed to support the claim that BEACON is a robust, transferable curation pipeline rather than a single tuned run on CHILDES.","section":"Section III-B/C, Section VI"}],"minor_comments":[{"comment":"Abstract and Table I: the ASR subset is variously “283-hour” / “282.6 h”; keep a single rounded figure consistently.","section":"Abstract, Table I"},{"comment":"Figure 1 is dense; the three-stage flow (inference → per-model alignment → ensemble voting) would be clearer with a short caption walk-through of one example utterance (e.g., “want to play”).","section":"Figure 1"},{"comment":"Eq. (6): the agreement rule mixes absolute endpoint tolerance and IoU-with-center; a one-sentence intuition for why both clauses are needed would help readers implement the consensus.","section":"Section III-C, Eq. (6)"},{"comment":"Table III: report absolute hours retained after the ins/del filter for FASA vs. BEACON side-by-side in the table header so yield differences are immediately visible when reading ΔWER.","section":"Table III"},{"comment":"Related Work: briefly note how BEACON differs from ROVER-style hypothesis combination (Fiscus 1997), which operates on word sequences rather than time-boundary votes, to avoid conflating the two ensemble traditions.","section":"Section II"}],"recommendation":"minor_revision","confidential_remarks":"Solid systems/data paper for eess.AS; the downstream Table III result is the load-bearing contribution and stands independently of the proxy’s absolute-accuracy gap. I would not block on human boundary annotation if the authors clearly reframe V-A, but a small annotated sample would raise the paper’s ceiling. Fit and novelty of the public CHILDES-aligned release are appropriate for the venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that they ship a public 413/283-hour CHILDES-derived resource with corrected utterance timestamps and show clear out-of-domain WER drops on four held-out child benchmarks across three model families. That is the result that matters.\n\nWhat is new is the specific pipeline: align each of four diverse ASR word streams (Parakeet TDT, Canary, WhisperX, Qwen3) to the trusted CHAT transcript via chunked edit-distance windows + soft monotonic DP, then majority-vote the spans with VERIFIED/RECOVERED/UNRESOLVED labels. Forced alignment, ROVER-style voting, and long-form tools (BatchAlign2, FASA, SailAlign) already exist; the contribution is the combination that keeps speaker-attributed CHAT turns instead of discarding them. They release code and both dataset versions, so it is checkable.\n\nThey do the evaluation right for the headline claim. Table III is zero-shot vs FASA-curated vs their 282.6 h set on RSR/MyST/OGI/CMU with Whisper, Parakeet, and Canary; relative average reductions of 19.5/3.6/5.6 % are real and not circular. The independent Granite proxy and the ins/del ≤ 0.25 filter keep the training pairs honest enough that the downstream numbers stand on their own.\n\nSoft spots are real but secondary. Timestamp quality is measured only by bounded WER of an external ASR on the cut clips; there is no absolute onset/offset error against human times and no hyper-parameter ablation (dozens of thresholds set by light inspection). Abstention can introduce selection bias. Section VI already flags both. These limit any strong claim that the boundaries are “accurate,” but they do not reverse the fact that the filtered clips improve three ASR families on non-overlapping data. Citation pattern is clean; related work covers the right prior art.\n\nThis is for people who train or evaluate child ASR, or who need a reusable recipe for long-form transcript-to-audio alignment when timestamps are junk. It is not a theory paper. I would send it to peer review; the data release and multi-model gains are enough to deserve referee time. Engage if you work on child speech or dataset curation; the resource is immediately usable.","headline":"Solid engineering release: multi-model consensus turns noisy CHILDES into a usable 283 h child-ASR set with real held-out gains; proxy-only timestamp eval is the main soft spot, not a load-bearing flaw.","tokens_in":16581,"tokens_out":593,"would_cite":true,"duration_ms":5702,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Multi-model ASR consensus recovers usable utterance timestamps from long, noisy child-speech recordings and turns CHILDES into training data that cuts out-of-domain child WER by up to 19.5%.","keywords":["child speech","automatic speech recognition","timestamp curation","ensemble alignment","CHILDES","long-form audio","forced alignment","dataset release"],"falsifier":"Manually time-align a representative sample of CHILDES utterances to gold onset/offset labels and check whether BEACON's consensus boundaries show lower absolute timing error than the original CHILDES timestamps, BatchAlign2, and FASA on that same sample.","tokens_in":16497,"feed_emoji":"🗣️","tokens_out":689,"duration_ms":5874,"temperature":0.7,"pith_summary":"Long naturalistic child-adult recordings in CHILDES come with trusted human transcripts but often with noisy, missing, or misaligned utterance timestamps, so the audio cannot be cut into clean clips for speech-model training. The authors introduce BEACON, a fully automatic pipeline that runs several diverse off-the-shelf ASR systems, aligns each system's word-level timestamps back to the trusted transcript, and fuses the resulting candidate boundaries by majority consensus (or abstains when models disagree). Applied to English CHILDES, the pipeline yields a 413-hour general-purpose release that keeps the original raw CHAT annotations and a stricter 283-hour ASR-ready subset with verbatim-normalized transcripts and extra audio-text agreement filters. Fine-tuning three different ASR families on the ASR subset improves word-error rates on four held-out child-speech benchmarks that never appear in the training data, with the largest average relative reduction reaching 19.5%. The same recipe is presented as corpus-agnostic: any long recording paired with a trusted transcript can be timestamp-curated without new human annotation.","feed_headline":"Multi-model consensus turns noisy CHILDES into child ASR data","feed_subtitle":"A 283-hour curated subset cuts out-of-domain child word-error rates by up to 19.5 percent","key_machinery":"BEACON (Boundary Estimation via Alignment CONsensus): multi-model inference of word-level timestamps, per-model alignment of those timestamps to the reference utterance sequence via search-window determination, candidate generation, and monotonic DP path selection, followed by majority-consensus fusion (or abstention) of the per-model time intervals.","core_discovery":"BEACON recovers reliable utterance onsets and offsets for long-form child speech by aligning multiple off-the-shelf ASR word streams to a trusted CHAT transcript and taking a consensus vote over the resulting candidate spans; the resulting curated CHILDES clips, once quality-filtered, supply transferable supervision that lowers out-of-domain child ASR error.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Multi-ASR consensus fixes CHILDES timestamps for child speech","BEACON curates 413h child speech via ensemble timestamp voting","Consensus ASR boundaries turn noisy CHILDES into ASR training data","Aligned multi-model timestamps yield 19.5% lower child ASR WER","Ensemble voting recovers utterance spans from long CHILDES audio"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that lower word-error rate of a fixed external ASR on the cut clips is a faithful proxy for timestamp accuracy, without any direct measurement of onset or offset error against human ground-truth times.","fun_headline_variants_meta":{"raw":{"variants":["Multi-ASR consensus fixes CHILDES timestamps for child speech","BEACON curates 413h child speech via ensemble timestamp voting","Consensus ASR boundaries turn noisy CHILDES into ASR training data","Aligned multi-model timestamps yield 19.5% lower child ASR WER","Ensemble voting recovers utterance spans from long CHILDES audio"]},"model":"grok-4.5","effort":"low","cost_usd":0.004414,"raw_usage":{"total_tokens":1339,"prompt_tokens":809,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":44140000,"prompt_tokens_details":{"text_tokens":809,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":437,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":809,"tokens_out":93,"duration_ms":3890,"temperature":1.0,"reasoning_tokens":437,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:47:33.234876+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Manually time-align a representative sample of CHILDES utterances to gold onset/offset labels and check whether BEACON's consensus boundaries show lower absolute timing error than the original CHILDES timestamps, BatchAlign2, and FASA on that same sample.","supporting_citations":[],"review_version":1}