MedASR BenchViMedCSS v1

Leaderboard Compare Datasets Methodology

normalizer v0.1

Leaderboard Compare Datasets Methodology

Loading the leaderboard…

MedASR Bench

Part of the Clinical Scribe program — Vietnamese medical ASR, measured in the open.

Leaderboard
Compare
Datasets
Methodology
Cite · arXiv:2602.12911
HF · tensorxt/ViMedCSS

Normalizer: normalizer v0.1
Harness: medasr-bench v0.4
License: benchmark data CC-BY-4.0
API snapshot policy: API results are date-stamped snapshots

03 · Datasets

One instrument, many corpora

MedASR Bench is a benchmark hub for Vietnamese medical ASR. Every dataset mounts into the same harness — same normalizer, same six metrics, same provenance rules — so scores stay comparable across corpora. 1 live today, 2 onboarding.

ViMedCSSVietMedsoonPrivate Clinic Setsoon

LiveAnchor datasetCC-BY-4.0

ViMedCSSVietnamese Medical Code-Switching Speech

Code-switched medical speech from Vietnamese lectures — every utterance carries at least one English clinical term.

Source · Vietnamese medical YouTube lectures

arXiv:2602.12911 HF · tensorxt/ViMedCSS

Dataset profile Leaderboard · 17 systems

Audio: 34.57h
Utterances: 16,576
CS terms: 889

Frozen splits · share of audio

Train: 24.31 h
Valid: 3.56 h
Test: 3.38 h
Hard: 1.38 h

Onboardingsoon

VietMed

Real-world Vietnamese medical speech recorded in actual doctor–patient conversations.

Source · Real-world medical speech recordings

What onboarding involves

OnboardingETA 2026

Private Clinic Set

Private conversational recordings from partner clinics, held out from public release to resist test-set contamination.

Source · Partner clinic consultations

What onboarding involves

Onboard a dataset

Run your corpus on the same instrument

Any Vietnamese medical speech corpus that meets six criteria can mount into the harness, and its scores are comparable from the first run. Datasets are frozen, versioned snapshots — never moving targets.

Open or contracted license
Transparent provenance
Fixed frozen splits
Contamination policy
Normalizer compatibility
Six-metric coverage

Criteria & versioning policy