PaperlayerView exact revision on arXiv

Source-linked AI summary

Leveraging Fine-grained Error Correction in Korean Speech Recognition for Consultation Services

Yonghyun Jun, Jimin Lee, Hwan Chang, Dongho Shin, Seolah Kim, Hwanhee Lee

arXiv:2609.09889v1cs.CL

TL;DR

Korean call-center ASR correction lacks realistic dialogue-level resources, especially when privacy constraints prevent access to audio. The paper introduces DasanCallDial and DCSC, combining token detection, gated routing, span correction, and dialogue context. DCSC achieves state-of-the-art performance, with pkoT5-based results improving detection accuracy and both balanced and erroneous-utterance WER over baselines.

  • Problem

    Korean text-only ASR post-editing lacks realistic dialogue-level resources and tailored methods, while privacy restrictions can prevent access to raw audio.

  • Method

    DCSC combines token-level detection, detector-guided routing, span-level correction, and dialogue-context augmentation for error-sparse Korean ASR transcripts.

  • Results

    90.43 detection accuracy, Bal-WER reduced from 14.67 to 13.30, and E-WER reduced from 29.35 to 26.27 are reported for pkoT5-based DCSC.

  • Takeaways & Limitations

    DasanCallDial provides a realistic benchmark and DCSC provides a strong baseline for Korean post-editing in text-only, error-sparse settings.

  • Takeaways & Limitations

    The study focuses on Korean text-only post-editing and leaves evaluation in other low-resource languages and multimodal audio-access settings for future work.

Abstract

from arXiv · show

Automatic Speech Recognition (ASR) technology is fundamental to customer service automation and large-scale transcription. However, even advanced ASR models exhibit inevitable errors in complex real-world environments such as call center conversations. When privacy restrictions preclude audio access, error correction must rely on text-based post-editing. Existing text-only approaches face significant challenges in low-resource languages, mainly due to a critical scarcity of annotated corpora and tailored correction methodologies. For Korean, this resource gap is particularly pronounced, as existing resources are predominantly designed for ASR training rather than text-based error correction. To address this, we introduce DasanCallDial, the first large-scale Korean benchmark dataset specifically curated for dialogue-level ASR error correction. Derived from genuine call center interactions, it comprises 1,974 dialogues with 115,460 utterances. Leveraging this resource, we propose Detector-Gated Contextual Span Correction (DCSC), a text-only post-editing framework for error-sparse Korean speech recognition transcripts. DCSC combines an encoder-based detector that first performs token-level error detection, followed by a language model-based corrector trained to rectify fine-grained span-level errors. Additionally, we employ dialogue-level context augmentation to enable the model to leverage discourse history for disambiguation. By employing multi-level granularity, our method achieves state-of-the-art performance, effectively overcoming the limitations of general LLMs in low-resource settings.

1 Introduction

Korean call-center ASR post-editing is hindered by noisy real-world speech, limited text-only resources, and insufficient discourse-level correction data. The paper addresses these gaps with DasanCallDial and DCSC, a context-aware multi-granularity correction framework.

  • Call-center ASR remains error-prone because of pronunciation variance, background noise, overlapping speech, and domain-specific jargon.
  • Privacy restrictions often block raw-audio access, making text-only transcript post-editing particularly valuable for public-service call logs.
  • DasanCallDial contains approximately 1,974 dialogue pairs and 115,460 utterance lines, with a 17.95% error rate across distinct utterances.
  • DCSC detects token-level errors, routes utterances to a corrector, edits fine-grained spans, and augments targets with preceding dialogue context.
  • pkoT5-based DCSC reaches 90.43 detection accuracy, reduces Bal-WER from 14.67 to 13.30, and lowers E-WER from 29.35 to 26.27.

2 Related work

Prior Korean resources largely target ASR training or synthetic, sentence-independent correction, leaving a need for realistic dialogue-level post-editing data. Related work motivates text-only detection, staged correction, and fine-grained supervision for low-resource settings.

  • KsponSpeech and ClovaCall provide Korean speech resources but omit raw ASR outputs, limiting their suitability for text-based error-correction research.
  • Existing Korean post-editing resources are limited by sentence-level independence, artificial noise or speech generation, and erroneous-only source samples.
  • DasanCallDial provides 115k parallel STT-and-ground-truth samples from actual call-center complaints while preserving a clean-majority, error-sparse distribution.
  • Text-only ASR error detection is motivated by settings where raw audio is unavailable, extending prior work beyond acoustic-transcript consistency models.
  • Two-stage pipelines that separate error localization from rewriting and fine-grained supervision improve correction performance under data scarcity.
  • Low-resource post-editing studies indicate that problem-specific inductive bias and domain-adaptive training can matter more than model scaling alone.

3 Dataset

DasanCallDial is constructed from anonymized real call-center recordings, with expert correction and validation producing dialogue- and utterance-level benchmark configurations. Its error analysis shows realistic speaker, topic, lexical, phonetic, and pronunciation variation.

  • Dataset construction: 296,100 seconds of recordings from approximately 370 counselors were transcribed and manually corrected into corresponding dialogue references.The recordings average 2 minutes and 30 seconds per call and total 82 hours and 15 minutes.
  • Dataset construction: Two experienced call-center personnel performed annotation and anonymization, followed by consensus discussion for disagreements.The annotation process involved independent expert work and later validation.
  • Dataset construction: Cohen’s κ was 0.73, while exact-match agreement between jointly identified erroneous utterances was approximately 85.5%.The double-annotation sample contained 1,000 test utterances.

4 Methodology

DCSC is a gated, two-stage text-only post-editing framework that detects token-level ASR errors before applying span-level corrections. It preserves full utterance context, augments inputs with preceding dialogue, and trains the corrector on error-focused examples to limit unnecessary edits.

  • Detector: DCSC first uses an encoder-based detector to identify token-level errors and decide whether an utterance requires correction.The detector produces a binary mask over tokenized transcription tokens.
  • Detector: The detector labels substitutions, deletions, and insertions as erroneous tokens, masking the preceding or following token for insertion cases.Matched tokens receive 0, while erroneous tokens receive 1 in the mask sequence.
  • Span-level correction: The corrector converts erroneous word-level regions into span pairs linking noisy text to corrected text, separated by [SEP], or outputs “No Error” for clean utterances.Consecutive erroneous tokens are merged into spans before constructing the correction target.
  • Dialogue-level context augmentation: Dialogue augmentation concatenates up to 10 preceding ground-truth utterances with the target utterance to resolve errors requiring discourse context.Special separator tokens distinguish the dialogue context from the target utterance.
  • Gated inference: At inference, the detector gates the corrector: clean utterances are returned unchanged, while flagged utterances are corrected using the full utterance rather than only predicted error tokens.The corrector receives dialogue context when context augmentation is enabled.
  • Gated inference: Corrector training combines genuinely erroneous samples, detector false positives, and a balanced random sample of error-free utterances to reduce spurious downstream corrections.The curated set is fine-tuned with the fine-grained span-level correction scheme.

5 Experiments

DCSC consistently outperforms passive and active baselines by combining stronger error detection with targeted correction that preserves clean utterances. Across model comparisons and ablations, pkoT5 with token-level supervision and span-level correction achieves the strongest overall correction results.

  • Baseline comparison: Direct utterance-to-utterance generation over-corrects clean utterances, whereas explicit detection gating better preserves already-correct transcripts.The pkoT5+Uttr2Uttr baseline reaches Bal-WER 25.33 because its N-WER remains high.
  • Performance overview: DCSC reduces E-WER from 29.35 to 26.27 and Bal-WER from 14.67 to 13.30 while maintaining low N-WER between 0.23 and 0.43.All DCSC configurations outperform the Zero-rule baseline on detection and correction quality.
  • K-Seq2Seq vs. LLMs: LLMs provide strong detection with dialogue context, but pkoT5 produces the lowest E-WER and Bal-WER, indicating more stable rewriting after correction is triggered.Qwen2.5 reaches Recall 75.32, while pkoT5 reaches F1 72.91 and the strongest correction metrics.
  • Comparison within K-Seq2Seq models: Among encoder–decoder models, pkoT5 balances Precision and Recall better than mT5 and KoBart, achieving F1 72.91, EM 85.16, and Bal-WER 13.30.mT5 and KoBart have higher Precision but substantially lower Recall, missing more erroneous utterances.
  • Model selection: pkoT5 achieves the best overall correction performance, with EM 85.16, E-WER 26.27, and Bal-WER 13.30, while requiring less computation than substantially larger LLMs.Its detection F1 is 72.91, higher than Qwen2.5 despite the smaller model size.
  • Effect of token-level detection: Token-level supervision improves KoElectra’s Recall from 76.11 to 81.19 and F1 from 69.85 to 70.61, with the largest gains from fine-grained error localization.The effect is not uniform across aggregate metrics or detector backbones.
  • Supervision granularity and context: Span-level targets reduce harmful rewriting and instability, while dialogue context further improves error identification and correction of context-dependent errors.For Llama-3.1, context raises F1 from 42.93 to 54.00 and lowers E-WER from 29.69 ± 0.62 to 26.62 ± 1.62.

6 Analyses

The analyses examine DCSC across transfer settings, speaker roles, error-span densities, and detector operating points. Results show strong cross-domain gains, role- and density-dependent behavior, and a tunable precision–recall trade-off.

  • Detector–corrector interaction: The corrector reduces false positives from 911 to 436, improving Precision from 62.46 to 74.64 and decreasing N-WER by 0.16 points.It recovers 475 of 911 falsely forwarded clean samples unchanged, but Recall falls from 81.20 to 68.72 after unresolved corrections.
  • Detector operating points: At the default detector threshold of 0.5, Recall is 81.20%, with 351 of 1867 erroneous utterances filtered out before correction.Lowering the threshold to 0.1 raises Recall to approximately 94.5%, while aggressive routing lowers standalone Precision.
  • Context prediction: Predicted context changes detection little but increases Bal-WER by 0.45 points for Llama-3.1 and 0.29 points for pkoT5.The corresponding Bal-Acc reductions are 0.45 and 0.01 points, indicating measurable but limited error propagation under predicted context.
  • Cross-domain transfer: Out-of-domain DCSC reduces Bal-WER from 22.66 to 18.19 with Llama-3.1 and 18.15 with pkoT5 without Hyper-BTS training data.Mixed-domain DSC further reaches 12.91 and 11.36, approaching in-domain SFT results of 11.08 and 10.52, respectively.
  • Performance by speaker role: Customer utterances achieve higher Bal-Acc than counselor utterances, 76.25 versus 75.14, while counselor utterances achieve lower Bal-WER, 10.81 versus 15.83.The analysis associates this pattern with customer pronunciation variability and counselors’ more structured, less error-prone speech.

7 Discussion

The discussion identifies scope boundaries involving domain generalization, interface granularity, under-correction, context coverage, metadata, and data governance. It also describes practical implications for audio-inaccessible transcription workflows.

  • Scope and modality: The text-only setting is designed for cases where raw audio is inaccessible, while multimodal approaches remain a future direction when audio is available.The study primarily focuses on Korean ASR transcript post-editing, leaving extension to other languages and tasks for future evaluation.
  • Dataset generalizability: DasanCallDial may not generalize perfectly to open-domain conversations because it reflects call-center protocols and frequent turntaking.The dataset’s topic and frequency diversity nevertheless indicates some task transferability.
  • Industrial impact: The framework provides a practical blueprint for audio-free post-editing in customer service systems, CRM analysis, and other accuracy-sensitive applications.This implication is framed for real-world, high-noise environments with resource constraints.
  • Interface granularity: The detector–corrector interface remains coarse because token-level detector masks still trigger correction at the utterance level.Passing token indices directly could enable more targeted edits but risks detection-error propagation, boundary sensitivity, and optimization instability.
  • Under-correction: Under-correction arises when the detector filters erroneous utterances or when the corrector fails to resolve forwarded errors.Lowering the detector threshold can reduce the first source but does not address corrector-side failures, motivating better calibration and training objectives.
  • Contextual limitations: Finer granularity can reduce immediate context and limit correction of errors requiring long-term dialogue cues outside the input window.Compressed dialogue-context embeddings or more sophisticated context augmentation are proposed to capture broader discourse dependencies.
  • Metadata and personalization: The dataset lacks explicit speaker metadata such as regional background or age group, limiting opportunities for personalized or dialect-aware correction.Future integration of such metadata must continue to follow privacy protocols.
  • Data governance: The dataset’s anonymization and ground-truth annotation were handled securely by trained Call Foundation employees with direct audio access.Only the anonymized dataset was shared with Chung-Ang University.

8 Conclusion

The work introduces DasanCallDial, a large-scale Korean benchmark for real-world ASR error correction, and DCSC, a two-stage framework for correcting Korean ASR outputs. DCSC combines detection, routing, dialogue context, and span-level correction, achieving state-of-the-art performance against zero-rule baselines and general-purpose LLMs.

  • DasanCallDial contains 1,974 call-center dialogues and 115,460 utterance lines with manually verified references.
  • DCSC combines token-level detection, detector-guided routing, dialogue-context augmentation, and span-level correction in a two-stage framework.
  • The primary KoElectra–pkoT5 model achieves state-of-the-art performance, outperforming zero-rule baselines and general-purpose LLMs in a label-imbalanced setting.The experiments also examine learning dynamics, transferability, and robustness.

CRediT authorship contribution statement

The contribution statement assigns authorship roles across writing, methodology, visualization, validation, conceptualization, data curation, supervision, project administration, and funding acquisition.

  • Yonghyun Jun contributed to writing, visualization, validation, methodology, investigation, formal analysis, and conceptualization.
  • Jimin Lee and Hwan Chang contributed to writing, methodology, and conceptualization, with Hwan Chang also contributing to validation and review and editing.
  • Dongho Shin and Seolah Kim contributed to data curation, while Hwanhee Lee contributed to writing, supervision, project administration, funding acquisition, and conceptualization.

Declaration of Generative AI and AI-assisted technologies in the writing process

The authors report using Claude-Opus-4.8 and GPT-5.6-Sol to improve readability during preparation of the work. They state that they reviewed and edited the resulting content and retained full responsibility for the publication.

  • Claude-Opus-4.8 and GPT-5.6-Sol were used to improve readability during preparation of the work.
  • The authors reviewed and edited the content after using the tools or services.
  • The authors state that they take full responsibility for the publication’s content.

Appendix A Dataset examples

The appendix presents examples of the original dialogue-level dataset alongside its utterance-level version. The examples illustrate dataset structure, typical content, and de-identification outcomes.

  • Figure 12 presents concrete examples from the original dialogue-level dataset.
  • Figure 13 presents examples from the utterance-level version of the dataset.
  • Side-by-side tables illustrate dataset structure, typical content, and the outcome of de-identification.

Appendix B Annotation guideline

Section 3.1.1 provides the detailed annotation guideline in Table 9.

  • Table 9 contains the detailed annotation guideline for Section 3.1.1.
  • The annotation guideline is presented as part of Section 3.1.1.
  • Table 9 is the referenced source for the section’s annotation details.

Appendix C Cross-ASR error distribution analysis

A Korean-fine-tuned Whisper transcription provides a cross-ASR check of DasanCallDial’s error-sparse characteristics. Whisper has a moderately higher error rate than HAIV, but the dominant error patterns remain closely aligned.

  • 20.90% utterance error rate for Whisper exceeds HAIV’s 17.95%, while approximately four out of five Whisper utterances remain error-free.
  • Noun errors dominate for Whisper and HAIV, at 50.2% and 56.3%, respectively.
  • Consonant-related errors remain dominant, comprising 82.6% for Whisper versus 84.5% for HAIV.
  • Omission and insertion errors are most frequent for Whisper and HAIV, at 80.9% and 75.2%, respectively.
  • The small error-rate gap and consistent dominant patterns suggest the error-sparse setting is not unique to HAIV.

Appendix D Detailed experimental setup

This appendix documents the detailed experimental setup and includes dataset examples, error-distribution analysis, and backbone-specific configuration tables.

  • Tables 10 and 11 provide the detailed experimental setup for Section 5.1.1.
  • Table 10 summarizes backbone architectures and computational requirements.
  • Table 11 reports training and inference configurations for each model family.
  • Figures 12 and 13 present dialogue-level and utterance-level dataset examples, respectively.
  • Figure 14 shows error distributions from re-transcribing DasanCallDial audio with a Korean-fine-tuned Whisper model.
Loading 2609.09889v1…