DasanCallDial

DasanCallDial is the first large-scale Korean benchmark built specifically for dialogue-level ASR error correction. It contains 1,974 real civil-complaint call dialogues (115,460 utterances) placed to the 120 Dasan Call Foundation, Seoul's municipal civic-information hotline. Each utterance pairs the transcription produced by a production speech-recognition system with a human-verified ground truth.

Unlike datasets built by injecting synthetic noise, DasanCallDial preserves the error-sparse distribution of a real deployed ASR system: only 9.83 % of raw utterances contain an error. That sparsity, rather than error density, is the central challenge the benchmark poses.

Introduced in "Leveraging Fine-grained Error Correction in Korean Speech Recognition for Consultation Services" (Engineering Applications of Artificial Intelligence).

• Training and evaluation code: https://github.com/yonghyun010102/Dasan_Project

Loading

from datasets import load_dataset

ds = load_dataset("zgold5670/DasanCallDial")
print(ds["train"][0])

To use it with the reference implementation, place the three files at data/dasancalldial/ inside the code repository:

hf download zgold5670/DasanCallDial --repo-type dataset \
    --local-dir data/dasancalldial
python scripts/prepare_data.py

Data fields

field	type	description
KEY	string	Dialogue identifier, derived from the call's recording date and time. Every utterance of the same call shares it.
utt_index	int	Position of the utterance within its dialogue, starting at 0. Sorting by (KEY, utt_index) reconstructs the conversation.
rec_mode	string	Speaker: tx = counselor, rx = citizen (customer).
in_out	string	Call direction (i inbound / o outbound). This is a property of the call, not a speaker label.
STT인식문	string	ASR transcription — the model input.
정답문	string	Human-verified ground truth — the target.
오류 여부	int	1 if STT인식문 != 정답문, else 0. Derived, provided for convenience.

Both text columns are already segmented into utterances; no further sentence splitting is required.

Splits

Partitioning is done at the dialogue level with a 0.80 / 0.05 / 0.15 ratio, so no conversation is shared between subsets and dialogue context never leaks across the boundary.

split	dialogues	utterances	avg. turns	error rate
train	1,579	91,355	57.86	9.70 %
validation	98	5,418	55.29	8.77 %
test	297	18,687	62.92	10.81 %
total	1,974	115,460	58.49	9.83 %

The partition ships with the dataset, so published results do not depend on a locally drawn split.

Deduplicated view

Real call-center speech repeats heavily — greetings, confirmations, filler. Deduplicating on (STT인식문, 정답문) within each partition removes those repeats. Because most removed duplicates are short, easily recognised, error-free phrases, the error rate roughly doubles. This is the view the reference models are trained and scored on; scripts/prepare_data.py derives it.

split	samples	avg. chars	error rate
train	50,470	22.58	17.54 %
validation	3,157	22.35	15.01 %
test	11,088	21.84	18.22 %
total	64,715	22.44	17.54 %

Deduplicate per partition, not globally — a global pass would move utterances across the split boundary.

Speaker asymmetry

Measured on the deduplicated view:

speaker	utterances	error rate
counselor (tx)	33,537	13.21 %
citizen (rx)	31,178	22.19 %

Counselor speech is professionally articulated and comparatively standardised, so it transcribes more accurately. Citizen speech varies with the complaint and is often less precisely articulated. Any model or metric that ignores this asymmetry will look better on counselor turns than it performs in the field.

Privacy and de-identification

De-identification was performed at the source, by trained 120 Dasan Call Foundation employees with authorised audio access, before any data left the Foundation. Only the anonymised result was shared for research. Personal names and identifying details are replaced with masks: *** (30,040 utterances) and OOO (10,801 utterances).

A regular-expression sweep over all 115,460 rows, across both text columns, finds zero occurrences of:

• telephone numbers (dashed or plain mobile form)
• resident registration numbers
• digit runs of five or more characters
• email addresses, URLs, card-like number groups
• vehicle registration plates

Location references remain, but only at district (구, 1.46 % of rows) and neighbourhood (동, 0.80 %) granularity — the level a municipal complaint requires. There are no lot numbers, no building-and-unit numbers, and effectively no street addresses (2 rows).

If you nevertheless find residual personal information, please open a discussion on this repository so it can be corrected.

Intended use and limitations

Intended for non-commercial research on Korean ASR post-editing, error detection, and dialogue-level correction.

• Domain-specific. Municipal civic-complaint calls in Korean. Vocabulary skews toward parking enforcement, transportation, public health, permits and welfare. Results will not transfer unchanged to other domains or to other languages.
• One ASR system. Errors reflect the specific production model used at the Foundation. A different recogniser makes different mistakes.
• Text only. Audio is not included and cannot be released; methods must be text-based post-editing.
• Masks are content. *** and OOO appear in both input and target. Models will learn to reproduce them; a system that overwrites a mask with invented text is failing, and evaluation should account for that.
• Not a speech-recognition training set. It is built for correcting ASR output, not for training an ASR model.

License

Released under CC BY-NC 4.0 (Attribution–NonCommercial 4.0 International): non-commercial use only, with attribution.

Anything derived from this dataset — including model weights trained on it — remains subject to these non-commercial terms. The reference code is separately licensed MIT, but that license does not extend to this data or to models trained on it.

Data were provided by the 120 Dasan Call Foundation for academic research.

Citation

@article{jun2025dcsc,
  title   = {Leveraging Fine-grained Error Correction in Korean Speech Recognition
             for Consultation Services},
  author  = {Jun, Yonghyun and Lee, Jimin and Chang, Hwan and Shin, Dongho and
             Kim, Seolah and Lee, Hwanhee},
  journal = {Engineering Applications of Artificial Intelligence},
  year    = {2025}
}