TEA-ASR-1.1-mini · Taiwan Everyday Audio 🍵

TEA-ASR is an open, drop-in speech-recognition model purpose-built for Taiwan Mandarin. It turns real speech into natural Traditional Chinese with authentic Taiwan vocabulary, and it stays robust through the everyday Mandarin–English code-switching common in Taiwan. Adapted from the state-of-the-art Qwen3-ASR foundation and merged into a single self-contained checkpoint, TEA-ASR loads and runs exactly like stock Qwen3-ASR — no converters, no post-processing.

TEA-ASR-1.1-mini is the 780M compact model (best accuracy-per-parameter) of the second generation. For the 2B flagship, see JacobLinCool/TEA-ASR-1.1. Compared with the first-generation TEA-ASR-1-mini, this release substantially improves code-switching — ASCEND and CSZS drop by 1.29 and 0.70 points absolute — with CommonVoice roughly level.

What's new in 1.1-mini

• 🔀 Code-switch leap — ASCEND 12.49 → 11.20, CSZS 13.21 → 12.51; embedded English is transcribed, not translated.
• 🏷️ Format tags — trained with output-convention tags: an optional decoder-prefix control that biases toward verbatim English and a chosen numeral style (see Format tags).
• 🪶 Still a single drop-in checkpoint, < 10 hours of public training audio, no runtime post-processing.

Quick start

pip install qwen-asr

from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained("JacobLinCool/TEA-ASR-1.1-mini")
result = model.transcribe(audio="utterance.wav", language="Chinese")[0]
print(result.text)   # -> Traditional Chinese with Taiwan lexicon

Set language="Chinese" for Taiwan speech (recommended). You can also pass a context= string of hotwords (names, jargon) for contextual biasing, exactly as with the base Qwen3-ASR.

Benchmark results

Mixed Error Rate (MER%, lower is better), all numbers from a single self-measured run under one protocol (see Evaluation). Columns: the two TEA-ASR-1.1 models, the original (unadapted) Qwen3-ASR bases, and two references — Breeze-ASR-25 (a Taiwan-specialist ASR) and Whisper-large-v3. Bold = this model.

Benchmark	TEA-ASR-1.1	TEA-ASR-1.1-mini	Qwen3-ASR-1.7B	Qwen3-ASR-0.6B	Breeze-ASR-25	Whisper-large-v3
CommonVoice 19 (zh-TW)	3.58	5.12	3.90	5.79	8.03	10.17
ASCEND (zh-en)	9.60	11.20	10.57	12.54	17.53	19.61
CSZS (zh-en)	10.94	12.51	11.03	16.03	12.18	23.24
NTUML2021	6.67	7.53	10.12	11.03	7.50	9.68

Generational improvement — TEA-ASR-1.1-mini vs TEA-ASR-1-mini (780M, same protocol, lower is better):

Benchmark	TEA-ASR-1.1-mini	TEA-ASR-1-mini	Δ
CommonVoice 19 (zh-TW)	5.12	5.14	−0.02
ASCEND (zh-en)	11.20	12.49	−1.29
CSZS (zh-en)	12.51	13.21	−0.70
NTUML2021	7.53	7.37	+0.16

How to read this. 1.1-mini delivers most of the 2B flagship's quality at well under half the parameters (780M vs 2B) and leads every 0.6B-class system in the table. Against the first-generation mini it is a clear code-switch upgrade (ASCEND −1.29, CSZS −0.70), trading a small step back on the in-domain lecture set (NTUML2021 +0.16). The metric folds away script differences (see Evaluation), so it does not reflect the decisive practical change: TEA-ASR emits Traditional script and Taiwan vocabulary natively, whereas the base produces Simplified script.

Format tags

TEA-ASR-1.1-mini was trained with output-convention format tags — an optional prefix, in the same channel as the language hint, that steers formatting without changing the recognition:

• keep-en — transcribe embedded English verbatim (do not translate dense code-switch).
• digits / zh-num — force Arabic (123) or Chinese (一二三) numerals.

Plain decoding (no tag) works well by default; the tags are for callers who need a specific convention. Try them interactively in the Space.

Evaluation

• Metric — Mixed Error Rate (MER). Character Error Rate for Chinese and Word Error Rate for the English tokens, computed jointly per utterance and micro-averaged.
• Content fold (applied uniformly to every dataset and every system). Before scoring, both the reference and the hypothesis are normalized to a common form — converted to Simplified Chinese with OpenCC (t2s), lowercased, and stripped of punctuation. This isolates recognition from script style, so a Simplified-output model and a Traditional-output model (TEA-ASR) are compared fairly on content. (TEA-ASR's actual output is Traditional; the fold is only for scoring.)
• Decoding. TEA-ASR and Qwen3-ASR are decoded with language=Chinese; Whisper-large-v3 and Breeze-ASR-25 use their own automatic language detection. All systems are scored with the same code on the same public splits; we do not import numbers reported elsewhere.

Dataset	What it tests	Eval split (n)
CommonVoice 19 (zh-TW)	Read Taiwan-Mandarin speech	full test (5013)
ASCEND	Spontaneous Mandarin–English code-switch conversation	full test (1315)
CSZS (zh-en)	Zero-resource code-switch benchmark	full test (3176)
NTUML2021	Mandarin lecture speech (university ML course)	test[:2000]

• No train/test leakage. Fine-tuning used only the training pools, disjoint from every evaluation split: the NTUML2021 train split, the ASCEND train split, and a CommonVoice slice drawn from validated_without_test. Evaluation runs on the full, untouched test splits; CSZS is not used in training at all. Every number above is leak-free.

How it was built

• Base Qwen/Qwen3-ASR-0.6B (AuT audio encoder + Qwen3 decoder).
• Adaptation: a rank-16 decoder LoRA (plus a low-LR encoder LoRA) trained on under 10 hours of public audio (CommonVoice zh-TW, ASCEND, NTUML2021, and TaiMECS), with general + code-switch replay and English-preservation training so dense code-switch is transcribed, not translated, plus error-analysis-driven targeted supplements and the format-tag conditioning above.
• Localization: Traditional-script + Taiwan-lexicon output is rendered through the model's own tokenizer (baked once at build time); there is no post-processing at inference.
• Packaging: the adapter is merged into the base and the localized tokenizer is shipped with it, so the release is a single drop-in checkpoint that loads like stock Qwen3-ASR (decode verified bit-exact on 152k+ sequences).
• Decoding tip: pass language="Chinese" for Taiwan speech; this also prevents translation-style outputs on dense code-switch.

Limitations

• Compact-model trade-off: on the hardest code-switch sets and on in-domain lectures the 780M mini trails the 2B TEA-ASR-1.1; for the best accuracy prefer the flagship.
• Scope: validated on the Qwen3-ASR family; loads via the qwen-asr package exactly like the base.

Acknowledgements

• Base model: Qwen3-ASR by Alibaba Cloud (Apache-2.0; underlying weights remain subject to the Apache-2.0 license and its attribution/NOTICE terms).
• TaiMECS (CC-BY-4.0).
• Benchmarks: Common Voice (Mozilla), ASCEND (CAiRE), CSZS, NTU ML2021.

Citation

@misc{teaasr2026,
  title  = {Tokenizer-First Adaptation of Mandarin ASR to Taiwan Mandarin},
  author = {TEA-ASR contributors},
  year   = {2026},
  note   = {TEA-ASR (Taiwan Everyday Audio); adapted from Qwen3-ASR}
}

The TEA-ASR adaptation and this checkpoint are released under the MIT License.