capstone ai-internship neurology sources


title: “Best Medical Transcription (ASR) Models” — Trelis Research source_type: blog_post date_ingested: 2026-07-01 url: https://trelis.substack.com/p/best-medical-transcription-asr-models original_date: April 9, 2026 author: Ronan (Trelis Research) tags: [ASR, speech-recognition, medical-transcription, benchmarks, Whisper, Gemini, Scribe, Speechmatics, Trelis]

Summary

Trelis Research (April 9, 2026) benchmarks 16 proprietary and open-source ASR models across 3 custom medical benchmarks. Key finding: Google’s proprietary model and Eleven Labs’ Scribe v2 lead; Whisper Large V3 is the best general-purpose open-source model. Domain-specific fine-tuning (MultiMed ST) shows poor generalization to accented speech and recent terminology. All 3 datasets are public on Hugging Face under Trelis Medical.

The 3 Benchmarks

MultiMed Hard — 50 hardest rows from YouTube medical content (MultiMed dataset). Entity CER (character error rate on medical terms only, not full sentence). Difficulty selected via median CER across Whisper Large, Canary, and Voxtral Mini.

EKA Hard — 50 hardest rows from Indian-accented English (EKA dataset, medical students reading prepared scripts with pharmaceutical + Indian-market medical terms). Tests generalization beyond American English.

Medical Terms 2025 — 50 synthetic examples of terms newly appearing in 2025 FDA/EMA/WHO sources. Tests model knowledge of recent medical terminology (near/after many models’ training cutoffs). Generated via LLM → Kokoro TTS.

Model Performance

MultiMed Hard (YouTube medical content)

ModelEntity CERNotes
Scribe v2 (Eleven Labs)13.4%Best overall
Gemini 2.5 ProHigh
Ursa 2 Enhanced (Speechmatics)High
Whisper Large V3HighTop open-source
Nova V3High
Whisper Large TurboLower
Whisper SmallLower
Parakeet Universal 3 Pro (AssemblyAI)Lower
Canary (Nvidia)Lower
Whisper 3 (Fireworks)Lower
VoxtralLow
Med ASR (Google open source)LowNear Whisper Tiny
Whisper Tiny / BaseLowest

Key: MultiMed ST (fine-tuned on MultiMed training set) performed well on this benchmark (in-distribution) but dropped significantly on EKA Hard (out-of-distribution), scoring similarly to Whisper Small.

EKA Hard (Indian-accented English)

ModelNotes
Gemini 2.5 ProLowest entity CER
Scribe v2Close second
ParakeetNotably better here than on MultiMed
Speechmatics Ursa
Universal 3 Pro (AssemblyAI)
Nova 3
Whisper Large V3 / TurboVery close

Medical Terms 2025 (recent terminology)

ModelNotes
Gemini 2.5 ProBest
Scribe v2Close second
Universal 3 Pro
Whisper Large V3 (Fireworks + direct)
Canary Turbo

Speechmatics models ranked lower on this benchmark. Google open-source Med ASR performed below even Whisper Tiny — significantly underperforming Google’s proprietary model.

Key Findings

  1. Google proprietary >> Google open-source Med ASR. Google’s proprietary ASR leads; the open-source Med ASR ranks near the bottom, close to Whisper Tiny. This is a significant gap within the same organization.

  2. Domain-specific fine-tuning generalizes poorly. MultiMed ST (fine-tuned on MultiMed training data) scored well on in-distribution test data but fell to Whisper Small–level on Indian-accented speech and recent terms. Domain-specific training ≠ generalizable medical ASR.

  3. Whisper Large V3 is the best general open-source model. Despite being older, Whisper Large V3 generalizes well across all 3 medical benchmarks. A strong foundation for custom fine-tuning rather than training from scratch.

  4. Best models achieve 10–20% entity CER on hardest cases. On easy/general cases, error rates are much lower. These benchmarks specifically selected the hardest 50 rows — they differentiate models, not represent typical performance.

  5. Eleven Labs Scribe v2 is the top proprietary model for medical ASR. Scribe v2 leads MultiMed Hard at 13.4% entity CER and is competitive on the other two benchmarks.

  6. Accented speech remains a challenge. Gemini 2.5 Pro leads on EKA Hard (Indian-accented), demonstrating superior accent robustness.

How to Access

All 3 datasets + full prediction results per model per benchmark are on Hugging Face under the Trelis Medical collection: search “Trelis medical” on huggingface.co.

Connected Pages

Entity - Abridge, Source - Abridge AI Evaluation Whitepaper, Concept - AI Evaluation in Healthcare