woongstar/ko-finance-asr-corrections

ko-finance-asr-corrections

Frequency-annotated Korean ASR confusion pairs from finance/stock YouTube.

• —210 pairs <!-- stat:pair_count -->
• —mined from 2,391 videos <!-- stat:scanned_videos --> of auto-captions
• —across 47 channels <!-- stat:corpus_channels -->
• —totalling 1,080.1 hours <!-- stat:corpus_hours -->

Each pair carries how often the term was mangled and how often it was said correctly, plus verification provenance.

한국어 금융·주식 유튜브 자동자막에서 실측한 ASR 오인식→교정 쌍입니다. 모든 쌍에 오표기·정답 표기 빈도(→ 용어별 오인식률)와 검증 메타데이터(2-LLM 합의 감사, 승격 티어)가 붙어 있습니다.

What makes it different

No public "confusion pair dictionary with observed frequencies" exists in any language (as of a 2026-08 survey), and Korean ASR error-correction data is nearly nonexistent — the only prior public pair set is KEBAP (2,478 pairs, error-type labels, frozen since 2023) and no finance-domain resource exists. Frequency is what makes a pair list actionable: it says which errors dominate real traffic:

• —펀더멘탈 → 펀더멘털 ×1,151 <!-- stat:count:펀더멘탈 -->
• —하이니스 → 하이닉스 ×904 <!-- stat:count:하이니스 -->
• —변합기 → 변압기 ×682 <!-- stat:count:변합기 -->
• —FMC → FOMC ×1,106 <!-- stat:count:FMC -->

Those four were 449, 342, 333 and 270 one snapshot ago; the counts grow with the corpus, which is why every number in this card carries a marker the release checker verifies.

Fields

wrong, right, corpus_count, observed_count, tier, evidence, category, auditor_models, approved_at, word_boundary, apply_scope. Field-by-field definitions: `docs/SCHEMA.md`.

How it was verified

No gold labels exist for this task, so verification is a loop, not an annotation pass: two independent LLM auditors judge each candidate, only consensus is adopted, an external registry check confirms stock names, and pairs are promoted through tiers with instant rollback. Details and per-round numbers: `docs/METHODOLOGY.md`.

Error rates, not just frequencies

Each pair carries corpus_count (the misrecognized form) and right_count (the verified one) from the same scan, so a row states how often the term came out wrong: FMC → FOMC 74.8% <!-- stat:rate:FMC -->, 엔트로픽 → 앤트로픽 87.3% <!-- stat:rate:엔트로픽 -->. 23 pairs have right_count: 0 <!-- stat:rightcountzero --> — across all 2,391 videos <!-- stat:scanned_videos --> the captions never once produced the correct spelling. Stock rows also carry ticker and market, so they join to price data without matching on a name.

data/withdrawn.json lists pairs that shipped and were later removed, with how many of their matches were genuine and which narrower key replaced them. A withdrawal says the key was unsafe as written, not that every correction it made was wrong.

Limitations (read these)

One corpus, one ASR system, one domain; 210 pairs <!-- stat:paircount --> is a seed, not a lexicon; frequency-ranked mining skews toward what these channels talk about; `corpuscount` counts surface forms, not confirmed errors; some pairs encode a spelling standard rather than a mishearing; tier B pairs were verified by auditor consensus, not by a human; no person-name pairs and no source sentences ship, by policy. The full list is in the repository README — read it before using the data as a general Korean ASR error list, because it is not one.

Benchmark

A standard-library scorer, a 1,100-item <!-- stat:eval_items --> synthetic evaluation set, and three dictionary baselines live in the GitHub repository. Fixes and over-corrections are scored together, because a correction dictionary that fixes 100 errors while damaging 20 correct sentences is not a good dictionary.

License and citation

Data: CC BY 4.0. Code (benchmark, scripts): MIT. Cite the concept DOI `10.5281/zenodo.22700919`, which always resolves to the latest snapshot; machine-readable metadata is in CITATION.cff in the GitHub repository.

Produced by the 껄무새 pipeline.