README.md

---
title: WeeMed AI
emoji: 🩺
colorFrom: blue
colorTo: green
---

# WeeMed AI

**Building a globally standardized, AI-driven digital health and long-term care platform
for aging societies.**

FHIR-first. Integrated health data, intelligent decision support, personalized preventive
medicine — deployed in real clinics, health-screening centers, and community long-term
care sites in Taiwan.

> **AI accelerates digitization; it isn't the product.**

A spin-off of the **Medical Image Processing Laboratory (MIPL)**, National Yunlin
University of Science and Technology — a lab that has built AI-assisted diagnostic
systems with National Taiwan University Hospital Yunlin Branch, National Cheng Kung
University Hospital, and Dalin Tzu Chi Hospital since 2003.

---

## What we release here

We ship models that fall out of real deployment problems, and we publish the
**measurements**, not just the weights.

### 🎙️ [Breeze-ASR-26-edge](https://huggingface.co/weemed/Breeze-ASR-26-edge)

Taiwanese Hokkien (台語) + Mandarin speech recognition, quantized for the edge, in two
runtimes — with a measured benchmark telling you which to pick:

| Runtime | RTF (CPU, 4 threads) | Real-time? |
|---|---|---|
| CTranslate2 / faster-whisper | **0.18 – 0.23** | ✅ |
| ONNX / sherpa-onnx | 1.17 – 1.76 | ❌ (but runs on Android/iOS/WASM) |

Measured on real multi-speaker meeting audio, not synthetic clips. Derived from
[MediaTek-Research/Breeze-ASR-26](https://huggingface.co/MediaTek-Research/Breeze-ASR-26)
(Apache-2.0) — all credit for the model to MediaTek Research.

---

## Why this matters to us

Taiwan is aging fast, and the people doing the caring — nurses at screening centers,
care managers at community sites, elders themselves — mostly do not speak to each other
in written Mandarin. They speak **Taigi**, in noisy rooms, on tablets held in one hand.

Health tech that only understands clean written Mandarin does not meet them where they
are. So we work on the unglamorous end of clinical AI: the languages actually spoken,
the devices actually held, the constraints actually present.

---

## How we publish

- **Honest benchmarks.** We report the number that survives scrutiny, including when it
  kills our own hypothesis. Our ASR card documents the measurement traps we fell into
  before publishing (short clips and repetitive audio both give flattering, wrong RTFs).
- **Named gaps.** What we could not measure is written down as an open gap, not omitted.
- **Real audio, real clinics.** Benchmarks on synthetic data tell you about synthetic data.

---

<sub>Apache-2.0 unless stated otherwise · Made in Yunlin, Taiwan</sub>