Leveraging Fine-grained Error Correction in Korean Speech Recognition for Consultation Services
Yonghyun Jun, Jimin Lee, Hwan Chang, Dongho Shin, Seolah Kim, Hwanhee Lee
cs.CL
Quick verdict
Introduces Korean call-center ASR-correction benchmark (115k utterances) and DCSC; dataset is valuable but WER gains are modest and reference bias lingers.
Summary
The paper introduces DasanCallDial, a benchmark of 1,974 real call-center dialogues (115,460 utterances) from Seoul's Dasan Call Center, pairing production ASR transcripts with manually corrected references; 17.95% of deduplicated utterances contain errors, preserving the clean-majority distribution of deployed systems. It also proposes DCSC, a two-stage text-only post-editor: a KoElectra token-level detector gates routing to a corrector that produces span-level edits conditioned on dialogue context. On the test split, pkoT5-based DCSC reaches detection F1 72.91, EM 85.16, and reduces balanced WER from 14.67 (zero-rule) to 13.30 and E-WER from 29.35 to 26.27 with N-WER 0.35, outperforming adapted fuzzy-logic, character-level, and utterance-to-utterance baselines; Llama-3.1 (13.40) and Qwen2.5 (13.52) variants trail slightly. Ablations show each component contributes; three-seed Wilcoxon tests support primary gains. Analyses quantify detector false negatives (18.80%), corrector recovery of false positives (52.14%), threshold sensitivity, and small predicted-context degradation (Bal-WER +0.29 to +0.45). Limitations: references derive from manual corrections of the same ASR system's output (HAIV bias acknowledged in Appendix C), E-WER reduction is modest (~10% relative), end-to-end recall falls to 68.72 after correction, and several backbones are single-seed. Code and the CC BY-NC dataset are released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- DasanCallDial is the first large-scale, real-world, dialogue-level Korean benchmark for text-only ASR error correction, with an error-sparse clean-majority distribution (1,974 dialogues, 115,460 utterances, 17.95% utterance error rate after deduplication).
- Decoupling error localization from rewriting — token-level detection, span-level correction targets, and dialogue-context augmentation — outperforms direct utterance-to-utterance post-editing under error sparsity.
- pkoT5-based DCSC achieves the best results (detection F1 72.91, EM 85.16, E-WER 29.35→26.27, Bal-WER 14.67→13.30 vs zero-rule), beating adapted fuzzy-logic, character-level, and Uttr2Uttr baselines.
- A 220M Korean-specialized encoder-decoder (pkoT5) matches or beats 7-8B LLM correctors at far lower deployment cost.
- Effective post-editing is feasible text-only, without audio access, in real operational call-center conditions.
Strengths
First real-world, error-sparse Korean post-editing benchmark
1,974 dialogues / 115,460 utterances from a government call center with raw production-ASR output retained; the Table 1 contrast with KsponSpeech, ClovaCall, KEBAP and Hyper-BTS is well argued — the others lack real audio, STT transcripts, or the clean-majority distribution needed to evaluate error-sparse editing.
Exemplary component ablation
Section 5.3 isolates each contribution on pkoT5: Bal-WER 22.13 (Uttr-to-Uttr) → 15.54 (span targets) → 13.55 (+dialogue context) → 13.30 (+detector gating), with three-seed standard deviations and Wilcoxon tests on the final step.
Transparent pipeline forensics
Correct metrics for class imbalance
External-validity checks
Weaknesses
Modest headroom on the core correction metric
Reference-construction circularity
Uneven statistical treatment across backbones
Underspecified deduplication/rebalancing
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: DasanCallDial & DCSC (arXiv:2609.09889; EAAI 183, 2026)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DasanCallDial?
A Korean benchmark for text-only ASR error correction built from 1,974 real call-center dialogues (115,460 utterances) of Seoul's Dasan Call Center. Each production ASR transcript is paired with a human-corrected reference; after deduplication about 18% of utterances contain errors, preserving the realistic clean-majority distribution of deployed speech recognizers.
What is DCSC and how much does it improve transcripts?
DCSC (Detector-Gated Contextual Span Correction) is a two-stage post-editor: a KoElectra token-level detector decides whether an utterance needs correction, then a corrector (best: pkoT5) rewrites error spans using dialogue context. On its test set it cuts balanced WER from 14.67 to 13.30 and error-utterance WER from 29.35 to 26.27 while nearly leaving clean utterances untouched (N-WER 0.35).
Is the dataset or code publicly available?
Yes. The paper's data-availability statement says the code is on GitHub and the dataset on Hugging Face, released under CC BY-NC 4.0 (non-commercial research use). Full training hyperparameters are listed in Appendix D.
Why not just use a large LLM to fix ASR transcripts?
In this paper's error-sparse setting, fine-tuned direct utterance-to-utterance LLM editing catastrophically rewrote clean utterances — Llama-3.1 reached N-WER 83.29 and Bal-WER 63.69 versus the 14.67 zero-rule baseline — whereas the detector-gated span approach kept N-WER at or below 0.43, showing the gating stage is what prevents over-correction.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The excerpt contains two '[...sampling gap...]' omissions (one around mid-methods/5.1, one covering late Section 5.3 through 6.5-6.7 and part of the references); this review cites only tables and numbers fully visible in the provided text (Tables 1-8, 10-11, Figs. 6-8 prose).
- Whether zero-shot prompted-LLM baselines were run anywhere in the full paper could not be verified from the excerpt; the fully visible Table 4 lists no such row.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.