README.md

---
title: Kana Whisper — Japanese ASR
emoji: 🎙️
colorFrom: indigo
colorTo: pink
sdk: gradio
app_file: app.py
pinned: false
license: mit
short_description: Japanese speech → katakana with sbintuitions/kana-whisper
---

# 🎙️ Kana Whisper — Japanese ASR → Katakana

A Gradio demo that transcribes short Japanese audio clips into **katakana**
(pronunciation, not kanji) using
[`sbintuitions/kana-whisper`](https://huggingface.co/sbintuitions/kana-whisper)
— Whisper `large-v3-turbo` fine-tuned by SB Intuitions (~0.8B params, MIT).

Example output: `キョーワイイテンキデスネ`

## Deploying to Hugging Face Spaces (ZeroGPU)

1. Create a new **Gradio** Space and push these files (`app.py`,
   `requirements.txt`, `README.md`).
2. In the Space **Settings → Hardware**, select **ZeroGPU**. The
   `@spaces.GPU` decorator in `app.py` reserves an H200 slice per request and
   releases it when the call returns.
   (ZeroGPU requires a PRO account or a Space in an org/grant that allows it.)

That's it — `float16` inference is used per the model card's recommendation.

## Run locally

Requires a CUDA GPU (the app loads the model onto `cuda`):

```bash
pip install -r requirements.txt
python app.py
```