TEA-ASR-1.1-mini-fmt · Taiwan Everyday Audio 🍵 (format-controllable, 780M)

The output-convention-controllable variant of TEA-ASR-1.1-mini. Same compact drop-in Taiwan-Mandarin ASR — native Traditional Chinese + Taiwan vocabulary — plus a working numeral-style dial (digits → 108年 / zh-num → 一百零八年) in the decoder-prefix channel.

Which one should I use?

• Formatting-sensitive, Mandarin-dominant work (subtitles, records, logs with a required numeral convention) on a small compute budget → this model.
• Any Mandarin–English code-switching → use TEA-ASR-1.1-mini (or the 2B TEA-ASR-1.1): at 780M the format control costs a substantial amount of dense-code-switch robustness (CSZS 12.51 → 14.83; see the table). This variant is for Mandarin-dominant audio.

Format control

Same interface as TEA-ASR-1.1-fmt: force the decoder prefix language Chinese format digits<asr_text> / … format zh-num<asr_text>.

Measured control strength (multi-digit panels, scripts/probe_numeral_flip.py):

Panel (audio)	pair flip	digits compliance	zh-num compliance
CommonVoice zh-TW	0.73	0.77	0.93
NTUML2021	0.27	0.43	0.70

(Measured on this released checkpoint.)

At 780M the dial is a moderate bias: strong on read Taiwan-Mandarin speech, weaker on spontaneous lecture audio; single digits and decimals follow the domain's natural convention.

Benchmark results

MER% (lower is better), same protocol as the family cards (content fold, full test splits).

Benchmark	TEA-ASR-1.1-mini-fmt	TEA-ASR-1.1-mini	Qwen3-ASR-0.6B	Breeze-ASR-25
CommonVoice 19 (zh-TW)	5.27	5.12	5.79	8.03
ASCEND (zh-en)	11.25	11.20	12.54	17.53
CSZS (zh-en)	14.83	12.51	16.03	12.18
NTUML2021	7.44	7.53	11.03	7.50

How to read this. The fmt variant slightly leads on lectures and ties ASCEND, but pays a real dense-code-switch premium (CSZS +2.3) for the numeral dial — at 780M the conditioned output space is small enough that convention training measurably competes with code-switch robustness (the 2B fmt variant pays only +0.35 there). Use it where the numeral convention matters and the audio is Mandarin-dominant; otherwise prefer the main mini.

Evaluation, data, and packaging

Identical protocol and packaging to TEA-ASR-1.1-mini: leak-free train/test splits, < 10 hours of public training audio, rank-16 decoder LoRA + low-LR encoder LoRA merged into one drop-in checkpoint, Traditional output from the model's own tokenizer (no runtime post-processing; decode verified bit-exact on 152k+ sequences). The fmt recipe adds numeral-convention counterfactual pairs (same audio, both conventions, opposite tags) mined from the same corpora at zero extra audio budget.

Acknowledgements & license

Adapted from Qwen3-ASR (Apache-2.0); TaiMECS (CC-BY-4.0); benchmarks: Common Voice, ASCEND, CSZS, NTU ML2021. Released under the MIT License.

@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}
}