README.md

metadata

title: MedASR Bench
emoji: 🩺
colorFrom: indigo
colorTo: purple
sdk: docker
app_port: 3000
pinned: false
license: mit
short_description: Vietnamese medical ASR benchmark, code-switching aware

MedASR Bench

A public benchmark for Vietnamese medical speech recognition with English code-switching, anchored on the ViMedCSS dataset. Part of the Clinical Scribe program.

This Space is the display layer only — a stateless Next.js app that renders the leaderboard, per-model pages, and methodology. It runs on free CPU; no audio and no model inference ever happen here. Benchmark numbers are produced by the offline harness (benchmark/medasr_eval.py) and published as versioned, auditable run manifests.

Metrics

WER · CER · CS-WER (code-switched-term WER) · N-WER · MTR (medical-term recall), each with bootstrap 95% CIs under a pinned text normalizer.

Running locally

npm install
npm run dev        # http://localhost:3000

Build / deploy

This Space is a Docker SDK Space. The image builds the Next.js standalone output and serves it with node server.js as uid 1000 on port 3000.

docker build -t medasr-bench .
docker run --rm -p 3000:3000 medasr-bench