TEA-ASR-1.1-fmt

TEA-ASR-1.1-fmt · Taiwan Everyday Audio 🍵 (format-controllable)

TEA-ASR-1.1-fmt is the output-convention-controllable variant of TEA-ASR-1.1. Same drop-in Taiwan-Mandarin ASR — native Traditional Chinese, Taiwan vocabulary, Mandarin–English code-switch — plus a working numeral-style dial: a decoder-prefix format tag that switches the SAME audio between Arabic (digits → 108年) and Chinese (zh-num → 一百零八年) numeral renderings.

Which one should I use?

• Formatting-sensitive, Mandarin-dominant work (subtitles, meeting/agency records, call-center logs that must follow one numeral convention) → this model.
• Heavy Mandarin–English code-switching → use TEA-ASR-1.1, which posts the best code-switch error rates of the family (this variant trades a measured amount of dense-code-switch robustness for the format control — see the table).

Format control

The tag rides the same decoder-prefix channel as the language hint:

from qwen_asr import Qwen3ASRModel
m = Qwen3ASRModel.from_pretrained("JacobLinCool/TEA-ASR-1.1-fmt")

# plain (recommended default): natural convention per domain
m.transcribe(audio="utt.wav", language="Chinese")

# force a numeral convention via the forced decoder prefix:
#   language Chinese format digits<asr_text>   -> Arabic numerals (108年)
#   language Chinese format zh-num<asr_text>   -> Chinese numerals (一百零八年)
#   append keep-en to bias embedded English toward verbatim transcription

(The public transcribe(language=...) API validates the language string, so pass the full prefix through the forced-prefix path — see the demo Space source for a 20-line reference implementation.)

Measured control strength (multi-digit test panels, deterministic first-30 selection; scripts/probe_numeral_flip.py):

(Measured on this released checkpoint.)

The dial is a strong bias, not a hard switch: expect it to flip most multi-digit renderings and to leave single digits (是1 / 是一) and decimals/percentages to the domain's natural convention. keep-en biases embedded English toward verbatim output; plain decoding already preserves English well, so its per-utterance effect is modest.

Benchmark results

Mixed Error Rate (MER%, lower is better), same protocol as the TEA-ASR-1.1 card (content fold: OpenCC t2s + lowercase + punctuation strip; full test splits; single self-measured run).

How to read this. The fmt variant matches the flagship on lectures (best-in-family 6.57) and ASCEND, and pays a measured premium on CommonVoice (+0.38) and dense code-switch (CSZS +0.35) for the numeral dial — the training mass that makes the tag causal necessarily shifts the conditioned output space. If you don't need convention control, use TEA-ASR-1.1.

Evaluation, data, and packaging

Identical to TEA-ASR-1.1: same leak-free train/test protocol, < 10 hours of public training audio (CommonVoice zh-TW, ASCEND, NTUML2021, TaiMECS), rank-16 decoder LoRA + low-LR encoder LoRA merged into a single drop-in checkpoint, Traditional output rendered by the model's own tokenizer (no runtime post-processing; decode verified bit-exact on 152k+ sequences). The fmt recipe additionally trains numeral-convention counterfactual pairs — the same real audio supervised under both conventions with opposite tags — mined from the same public corpora (no extra audio budget).

Acknowledgements & license

Same as TEA-ASR-1.1: 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}
}