README.md

---
title: TEA-ASR Taiwan Mandarin
emoji: 🍵
colorFrom: green
colorTo: indigo
sdk: gradio
sdk_version: "5.49.1"
app_file: app.py
python_version: "3.12.12"
pinned: false
short_description: "Taiwan-Mandarin ASR: Traditional + zh-en code-switch"
models:
  - JacobLinCool/TEA-ASR-1.1
  - JacobLinCool/TEA-ASR-1.1-mini
  - JacobLinCool/TEA-ASR-1.1-fmt
  - JacobLinCool/TEA-ASR-1.1-mini-fmt
---

# TEA-ASR — Taiwan Mandarin ASR 🍵

Demo for **TEA-ASR (Taiwan Everyday Audio)**: Traditional-script + Taiwanese-lexicon ASR with robust
Mandarin–English code-switch, adapted from Qwen3-ASR with a tokenizer-first procedure. **No runtime post-processing** —
the Traditional/Taiwan decode is baked into the model's own fast tokenizer.

- **Models**: `TEA-ASR-1.1` (2B flagship, default) and `TEA-ASR-1.1-mini` (780M) for the best recognition,
  plus the format-controllable `TEA-ASR-1.1-fmt` / `TEA-ASR-1.1-mini-fmt` — all public, no token needed.
- **Hardware**: ZeroGPU (`@spaces.GPU`).
- Set the language hint to **Chinese** for Taiwan speech (best results).
- **Format tags** (`-fmt` models only): pick a **numeral style** (Arabic 123 / Chinese 一二三) to render the
  same audio in that convention — a verified flip control, strongest on multi-digit numbers — and tick
  **Keep English** to bias code-switch English toward verbatim transcription. These teacher-force a
  decoder-prefix convention tag; the main-line models ignore them. For heavy code-switch use the main line
  (best CSZS/ASCEND); for formatting-sensitive Mandarin-dominant audio use `-fmt` (see model cards for the
  measured trade-off).