Loading the leaderboard…
Loading the leaderboard…
03 · Dataset profile
Code-switched medical speech from Vietnamese lectures — every utterance carries at least one English clinical term. The anchor benchmark of MedASR Bench — its first live dataset.
03.1 · Specimen
Two transcript lines from the corpus — Vietnamese clinical speech with the code-switched English terms highlighted, exactly the spans CS-WER scores.
Bệnh nhân có tiền sử tăng huyết áp, chỉ định siêu âm Doppler động mạch cảnh.
Kết quả MRI cho thấy tổn thương chất trắng lan tỏa, cần theo dõi multiple sclerosis.
03.2 · Splits
Test is the primary leaderboard. Hard isolates rare and unseen terms — when a model ranks well on Test but slides on Hard, it memorized the vocabulary instead of learning it.
| Split | Hours | Utterances | Share of audio | Role |
|---|---|---|---|---|
| Train | 24.31 h | 11,833 | 74.5% | Adaptation corpus — Track C systems may train on this split and nothing else. |
| Valid | 3.56 h | 1,714 | 10.9% | Hyperparameter tuning and checkpoint selection. Never reported as a result. |
| TestPrimary | 3.38 h | 1,615 | 10.4% | The primary leaderboard — every headline number on the home page is scored here. |
| HardGeneralization | 1.38 h | 658 | 4.2% | Rare and unseen terms, disjoint from Train — the generalization leaderboard. |
03.3 · Topic distribution
Five clinical topic areas; bar lengths are relative to the largest topic.
03.4 · Construction
Five stages from raw crawl to verified labels. Stage three is why the contamination policy exists.
Vietnamese medical channels crawled and filtered down to clean lecture-style audio, then segmented into 2–21 second utterances.
Candidate English medical terms located against the 64,232-entry Meddict bilingual medical lexicon to anchor code-switched spans.
First-pass transcripts and span labels drafted by Gemini 2.5 Pro — the origin of the permanent contamination flag on Gemini-family rows.
Surface variants of the same term collapsed to canonical forms by edit-distance matching, yielding 889 distinct code-switched medical terms.
Annotators verified every label; inter-annotator agreement κ = 0.65. Verified — but Gemini-seeded. See the caution below.
03.5 · Validity
Two caveats bound what ViMedCSS results can and cannot claim.
03.6 · The hub
MedASR Bench is multi-dataset by construction — every corpus mounts into the same harness, normalizer, and provenance rules.
34.57 h Code-switched medical speech from Vietnamese lectures — every utterance carries at least one English clinical term.
Real-world Vietnamese medical speech recorded in actual doctor–patient conversations.
Private conversational recordings from partner clinics, held out from public release to resist test-set contamination.