acceptodds
Under review as a conference paper at ICLR 2027

VeriSpeech-Repair: Geometry-constrained local repair with clean-context audits

Abstract

A speech editor may recover a missing word while changing audio outside the intended edit. We study this tension with VeriSpeech-Repair: a fixed set of local speech patches, an explicit no-repair choice, and separate tests for transcript recovery and preservation outside the original corruption span. On controlled LibriSpeech corruptions, broad patch selection roughly halves automatic word error rate (WER) but reduces exact outside-span context by an absolute 22.1% among outputs that can be aligned to the original audio. A score-free geometry rule preserves exact context under local splicing and lowers WER by an absolute 2.8% on 782 sources after excluding earlier calibration sources. In a separate comparison on those sources, the configured systems differ in architecture, input information, and proposal count. dots.tts.edit achieves 52.259% automatic transcript pass on corrupted cases, compared with 28.261% for the risk controller. Only five of the editor’s 2,346 corrupted outputs align to the original time map; none has exact context, and the other 2,341 cannot be assessed by this fixed-map measure. Transcript recovery and fixed-map preservation therefore give different outcomes for these systems.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Speech repair, speech editing, localized speech synthesis, abstention, content preservation, speech evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_NtRvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper explores how to keep context unchanged in speech repair. The authors propose several manually designed rules to generate candidates, along with a "no repair" candidate. These candidates are then selected using a WER-based validation method. The authors also analyze whether the repair alters the context outside the damaged span. Experiments are primarily conducted on the LibriSpeech dataset. Results show that direct editing (dots.tts.edit) achieves better recovery results, i.e., a higher pass rate and lower WER, but the method based on geometric constraint rules achieves complete context alignment and, in most cases, preserves exact context. The authors also analyze the biases between different ASR evaluation models.

Critical strengths

The authors described their experimental setup in detail and provided an anonymous code repository, which may be helpful to the domain.

Critical weaknesses
  1. This article appears to merely describe how to keep context unchanged in speech repair task using manually defined rules, which is simply an engineering improvement rather than a scientific innovation. The method's effectiveness is quite limited and is compared to only one baseline (dots.tts.edit), so it is difficult to position this paper within related work.
  2. The presentation of the paper is poor and difficult to understand. Adding more pictures and specific examples to illustrate the method (such as the nine proposals in candidate generation) may be helpful for readers to understand this paper.
Questions for authors

Can geometry-based policy be combined with existing models like dots.tts.edit to improve context invariance without compromising recovery performance? It would be good if the geometry rules could be plug-and-play across all speech repair models to achieve context-invariant speech repair.

3 comments

The study evaluates a bounded post-generation decision rule over a fixed proposal family. VITS and SpeechT5 supply patches. The nine local proposals include the ASR word-time interval, duration-scaled target-difference intervals, two deletion-gap placements, and four deletion-boundary shifts. Each candidate is spliced without cross-fading; the controller retains no-repair when no proposal passes its rule (pp. 3, 5–6). These are controlled rules for stated spans, not an online natural-fault localizer.

The paper includes no-repair, several local selector policies, and one direct editor comparator. It does not compare a broad set of editing systems with shared input and proposal budgets. The experiments support a narrow assessment of automatic transcript recovery and fixed-map context under controlled LibriSpeech corruptions; they do not establish general repair quality.

The paper describes proposal families but reports no audio examples or case-level assets. The direct comparison does not apply geometry filtering to dots.tts.edit: that editor receives a word-difference instruction and returns one output, while the controller uses nine VITS proposals plus no-repair and known synthetic spans. The results do not establish plug-in compatibility or preserved recovery after integration (pp. 6–8).

Applying geometry to dots.tts.edit outputs would require mapping its output timing to the original corruption span and evaluating content and preservation under matched conditions. The current submission reports no such result.

Thank you for the response. However, the authors have only clarified the current scope and limitations of the work, without providing new experimental results or supporting evidence. Since my concerns remain unaddressed, I will maintain my score.

Thank you for identifying the missing empirical evidence. We added a paired analysis of the archived dots.tts.edit outputs against no-repair. This is a new analysis of existing outputs, not a new generation run. It covers 2,346 corrupted cases from 782 source utterances (three corruption conditions per source). We average conditions within source and use 10,000 source-utterance cluster-bootstrap resamples (seed 20260919) for 95% intervals.

Under Faster-Whisper large-v3-turbo, the editor-minus-no-repair absolute difference in exact-transcript pass rate is +37.85% (95% CI [+35.34%, +40.41%]); the absolute WER difference is -13.41% ([-14.30%, -12.50%]). Under Wav2Vec2 CTC, the corresponding differences are +38.49% ([+35.98%, +40.92%]) and -17.39% ([-18.25%, -16.54%]). The archived outputs therefore show higher automatic transcript recovery than leaving the corrupted input unchanged under both frozen evaluators. These results do not establish naturalness, speaker match, or performance beyond this controlled set.

The fixed-map context comparison has a different limit: only 5 of 2,346 editor outputs align to the original time map; none of those five preserves exact context, and the other 2,341 are unassessable by this measure. The local controller aligns all 2,346 cases and preserves exact context in 1,236, but it uses nine proposals plus no-repair and known synthetic spans, while the editor returns one output. This is not a compute-matched comparison and does not establish that geometry can be applied plug-and-play to the editor. The revised paper reports the paired recovery analysis and clarifies these comparison limits in the Abstract, Results, Discussion, Limitations, and Conclusion.

Reviewer_DMFXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies local speech repair using existing TTS models to generate replacement segments, then selects a repair or keeps the input based on ASR scores and edit boundaries. It evaluates transcript recovery and preservation outside the corrupted span.

Critical strengths
  1. The problem is relevant: correcting words should not change surrounding speech.
  2. The paper reports negative results and checks with different ASR models.
Critical weaknesses
  1. The writing is hard to follow. a lot of terms are introduced without explanation, for example in the abs: terms like “speech patches,” “no-repair choice,” “geometry,” and “exact outside-span context” appear before their meaning is made clear.
  2. The novelty is limited. Given known corruption boundaries, copying audio outside the edit already guarantees its preservation.
  3. The preservation metric cannot assess 2,341/2,346 corrupted outputs from the competing editor, so the comparison is not very informative.
  4. No listening test. WER and unchanged surrounding samples cannot tell whether the replacement sounds natural or matches the speaker.
  5. The content gate's improvement does not transfer to other ASR evaluators. The provided code also cannot reproduce the audio experiments.
Questions for authors

This paper feels mostly AI written to me

1 comment

Known-span geometry provides a construction guarantee under a controlled condition: when the candidate splice lies within the original synthetic corruption span, has the operator-specific duration, and assembly copies every off-span sample, the mapped context is equal by prefix–suffix identity (pp. 4–6). The clean reference enters only after decisions are frozen. This tests the local-splice implementation under known boundaries; it does not establish natural-failure localization or restoration quality.

For dots.tts.edit, 5 of 2,346 corrupted outputs align to the fixed original-span map. The other 2,341 are unaligned, so the map cannot measure their off-span audio. This leaves preservation unmeasured for those outputs; it does not show that their context changed (pp. 7–8, 15). The paper reports no listening evaluation or speaker-identity assessment, so WER and fixed-map context do not establish naturalness or speaker match (pp. 3–4, 9, 13).

The geometry-only selector improves automatic content over no-repair under two frozen post-selection decoders. The score-gated increment is near zero under large-v3-turbo and reverses under Wav2Vec 2.0 CTC. Selection remains tied to Whisper-small (pp. 6, 8–9, 13–14).

The paper's reproducibility statement says the anonymous archive has text-only decision helpers and a toy example, while audio, model weights, predictions, and the benchmark runner are absent. The headline Transformers Whisper-small runtime revision is also unavailable (pp. 6, 10, 13). The AI-use statement on p. 10 discloses help with writing polish, related-work search, analysis-code review, and discussion of recorded results. The current prose remains difficult to follow; the submission contains no readability revision.

Reviewer_s2Z2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Authors propose VeriSpeech-Repair - a model for managing local speech restoration based off available text transcript. They compare several implementations of the manager against LibriSpeech corruptions using dots.tts.edit as a baseline for comparison. Model sees strong performance on the curated selection, suggesting utility of approach.

Critical strengths

Proposed model is justified in approach and appears easily extensible to other combinations of backend models for local speech restoration. Use of public datasets for demonstration allows easy reproducibility of approach on future applications. Numbers suggest encouraging performance and potential to benefit community.

Critical weaknesses

Critical weaknesses of the paper are clarity, thoroughness of evaluation, and lack of proper discussion at end. I will go in order:

  • Clarity: there is lack of clear definition of technical implementations. For instance, the authors refer frequently to 'hard geometry' but it is difficult to understand from the text the exact application they are referring to. Fleurs is used in calibration but I'm unsure exactly what they are calibrating using the dataset. There is a general assumption of subfield specific knowledge in the verbage of the presentation that is not ideal for the general audience the conference is for. I suggest to authors to revise and clarify much of their technical terminology and presentation so even their colleagues outside local text restoration can expound on their findings. Particular notion can be brought to figure 1, which is simply a collection of boxes used to indicate a diagram of the methodology. Readers would benefit from actual exposition here.

Further, the arrangement of graphs bleeds into the main paper. It is difficult to tell which of the captions are part of the main argument or description of related figures.

Final, please note that the use of bold in tables typically suggests best performing model per metric. This is not the case in your presentation and is suggested to be revised to avoid accidentally misleading readers.

  • Thoroughness of evaluation: while the proposed manager is relatively strong as a performer, it only has a single model as a baseline to compare against. Given the 'previous works' section of the piece, it appears that there is a healthy community of the subfield being investigated. Are there not more models to compare against to provide a stronger argument for the model? Is librispeech the only dataset available to compare the strengths of the approach? The varied number results leaves some doubt to how effective the model will be in the long-tail.

  • Discussion: the conclusion of the work is used simply to restate the results. This is not the general purpose of a discussion section; it is a place to conjecture and start a discussion with the general research community about the significance of the work and its findings. It would be advised to further develop these sections to aid the downstream engagement of the work.

Given the limited comparative numbers of the approach and the lack of detail regarding significance, I am of the opinion that the work should not be accepted to the conference and suggest a rejection. For amelioration of this decision, I suggest a future draft contain further expository information of results and clarification of technical terms for a wider audience, alongside an additional evaluation set/baseline to properly situate the model against. On those conditions I would be more confident providing an acceptance.

Questions for authors

Can you clarify the purpose for calibration and your 'geometry' approach?

Why is there only one model for comparison against? Are there not other comparative models for the approach or do you see this instance as a wholly novel approach to the problem of local restoration? If the latter, can you develop this further for the audience?

1 comment

The geometry rule uses recorded synthetic-failure metadata. A candidate's repair interval [p,q) must lie inside the original corruption interval [b,e), and its output duration must match the operator-specific expected duration within two samples. Geometry-only ranks eligible proposals by outside-span displacement, duration error, and proposal label; it returns no-repair when no proposal qualifies (pp. 4–6). The clean waveform and clean-reference score do not enter selection.

The paper describes FLEURS as decoder calibration only: 128 source utterances expand to 512 four-variant examples, and no FLEURS repair result enters the reported outcomes. The PDF does not specify which decoder parameter is calibrated, so the calibration detail is not established in the submission.

The evidence is controlled LibriSpeech repair with known synthetic spans. The direct dots.tts.edit comparison is one comparator; the systems differ in input information, architecture, and number of proposals. FLEURS is calibration only, and natural-failure localization is untested.

Figure 1 and its caption are legible on PDF page 2. The Table 3 highlight emphasizes the Risk controller although it is not best on every metric, leaving the visual convention unclear. The current manuscript does not explain that emphasis.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.