README.md

---
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](https://huggingface.co/datasets/tensorxt/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

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

## Build / deploy

This Space is a Docker SDK Space. The image builds the Next.js
[`standalone`](https://nextjs.org/docs/app/api-reference/config/next-config-js/output)
output and serves it with `node server.js` as uid 1000 on port 3000.

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