model card: fresh-eval numbers + protocol
Browse files
README.md
CHANGED
|
@@ -34,8 +34,6 @@ A companion **TEA-ASR-1** shares the identical recipe — see [`JacobLinCool/TEA
|
|
| 34 |
- 🔀 **Code-switch robust** — handles natural zh-en mixing instead of translating Mandarin into English.
|
| 35 |
- 🧩 **Drop-in Qwen3-ASR compatible** — same loading and inference API as the base model; nothing else to install
|
| 36 |
or call.
|
| 37 |
-
- 🚫 **No runtime OpenCC, no string post-processing** — Traditional-Chinese output is baked into the model's own
|
| 38 |
-
tokenizer; inference is a pure model forward + decode.
|
| 39 |
- 🪶 **Lightweight adaptation** — a small decoder LoRA on a frozen audio encoder, trained on a few hours of public
|
| 40 |
audio, then merged for deployment.
|
| 41 |
|
|
|
|
| 34 |
- 🔀 **Code-switch robust** — handles natural zh-en mixing instead of translating Mandarin into English.
|
| 35 |
- 🧩 **Drop-in Qwen3-ASR compatible** — same loading and inference API as the base model; nothing else to install
|
| 36 |
or call.
|
|
|
|
|
|
|
| 37 |
- 🪶 **Lightweight adaptation** — a small decoder LoRA on a frozen audio encoder, trained on a few hours of public
|
| 38 |
audio, then merged for deployment.
|
| 39 |
|