dorajim15/TEA-ASR-1.1-transcribe-cpp

TEA-ASR-1.1 GGUF for transcribe.cpp

TEA-ASR-1.1 (Taiwan Mandarin, Traditional Chinese output) converted to GGUF for transcribe.cpp. Used by the VoiceInk zh-TW fork.

台灣華語語音辨識模型 TEA-ASR-1.1 的 transcribe.cpp 格式(GGUF),給 VoiceInk 繁中版用。

File	Quant	Size (bytes)	SHA256
TEA-ASR-1.1-Q8_0.gguf	Q8_0	2,184,763,808	2e4aeb45c9e7f421d99909e29e49a54e204d70ed0e77bebc3f18a135d78298d0

How it was made

1. Source: JacobLinCool/TEA-ASR-1.1 revision bda08df76d4fd6b487b4a1dd7f0bddf8541696f8.
2. TEA-ASR ships only tokenizer.json. tea_prepare.py splits it into the vocab.json / merges.txt / chat_template.json that transcribe.cpp's converter expects.
3. transcribe.cpp commit 856d7c10: scripts/convert-qwen3_asr.py to BF16, then transcribe-quantize --quant Q8_0.

The conversion is reproducible: running it again gives the same SHA256.

Notes

• TEA-ASR marks 1,161 tokens with Private Use Area characters (U+E000–U+F8FF). The original tokenizer.json decoder strips them, but GGUF keeps them, so remove U+E000–U+F8FF from the output.
• Pass language Chinese (VoiceInk language zh), as the TEA-ASR authors recommend.
• transcribe.cpp 856d7c10 caps Qwen3-ASR decoding at 256 tokens. Split audio into chunks of at most 30 seconds.
• Output has almost no punctuation. VoiceInk adds it afterwards with CT-Transformer.

License

MIT, following TEA-ASR-1.1. The base model Qwen3-ASR-1.7B is Apache-2.0. All credit for the model goes to the TEA-ASR authors. This repository only changes the file format.