README.md

metadata

license: mit
task_categories:
  - automatic-speech-recognition
language:
  - en
tags:
  - medical
  - asr
  - entity-cer
  - benchmark
size_categories:
  - n<1K

MultiMed Hard — Medical ASR Benchmark

Entity-aware medical ASR benchmark — 50 hard rows from medical lectures and interviews.

Prepared by Trelis Research. Watch more on Youtube or inquire about our custom voice AI (ASR/TTS) services here.

Source

Derived from leduckhai/MultiMed EN test split (4,751 rows, MIT license). YouTube medical channels — lectures, interviews, podcasts, documentaries. Transcripts are human-reviewed.

Preparation

1. Filter: audio ≥ 2s, ≤ 29s, text ≥ 20 chars
2. Casing filter: drop all-caps / all-lowercase rows
3. Whisper CER filter: drop rows with whisper-large-v3 CER > 10% (bad GT alignment)
4. Gemini Flash entity tagging (6 medical categories)
5. Keep rows with ≥ 1 entity (entity text ≥ 5 chars)
6. 3-model difficulty filter (whisper-large-v3, canary-1b-v2, Voxtral-Mini) with whisper-english normalization
7. Exclude median entity CER > 0.9
8. LLM validation (Gemini Flash): drop non-medical content, generic entities, GT typos
9. Top-50 by median entity CER

Entity categories

• drug — drug or medication names (brand or INN)
• condition — diagnoses, diseases, syndromes, disorders
• procedure — surgical, diagnostic, or therapeutic procedures
• anatomy — anatomical structures, organs, body regions
• biomarker — lab tests, biomarkers, genes, proteins, molecular markers
• organisation — hospitals, regulatory bodies, pharmaceutical companies

Columns

• audio — 16kHz WAV
• text — ground truth transcript (human-reviewed)
• entities — JSON array of tagged medical entities with text, category, char_start, char_end
• difficulty_rank — 1 = hardest
• median_entity_cer — median entity CER across 3 difficulty-filter models

Leaderboard (16 models, sorted by Entity CER)

#	Model	WER	CER	Entity CER	Results
1	scribe-v2	0.100	0.060	0.134	results
2	MultiMed-ST (whisper-small-en)	0.115	0.075	0.160	results
3	gemini-2.5-pro	0.105	0.062	0.167	results
4	ursa-2-enhanced	0.105	0.060	0.196	results
5	whisper-large-v3	0.085	0.052	0.197	results
6	nova-3	0.120	0.069	0.199	results
7	whisper-large-v3-turbo	0.093	0.056	0.218	results
8	whisper-small	0.133	0.075	0.228	results
9	parakeet-tdt-0.6b-v3	0.159	0.101	0.233	results
10	universal-3-pro	0.125	0.100	0.234	results
11	canary-1b-v2	0.150	0.093	0.255	results
12	whisper-v3 (fireworks)	0.130	0.090	0.261	results
13	Voxtral-Mini-3B-2507	0.109	0.075	0.273	results
14	medasr	0.251	0.145	0.278	results
15	whisper-tiny	0.236	0.144	0.360	results
16	whisper-base	0.221	0.156	0.379	results

Evaluated with Trelis Studio, whisper-english normalization.