funnyQZQ/undertext-models

UnderText models

UnderText 是一个在 Mac 本地运行的实时英文字幕和中文翻译工具。这个仓库存放它用到的五个模型的副本，这样下载不依赖原作者的仓库一直可用。同样的文件在 ModelScope（中国大陆下载快）和 Hugging Face 上各有一份。文件和原版完全一样，没有做任何修改。

Copies of the five models used by UnderText, a local real-time English subtitle and Chinese translation app for macOS. Keeping them here means downloads don't depend on the original repositories staying available. The same files are on ModelScope (fast in mainland China) and Hugging Face. They are unchanged from the original releases.

文件 / Files

文件 / File	用途 / Role	来源 / Source	原作者 / Original authors	许可证 / License
Qwen3.5-2B-Q4_K_M.gguf	翻译 / Translation	unsloth/Qwen3.5-2B-GGUF	Qwen Team, Alibaba (Qwen/Qwen3.5-2B); GGUF quantization by Unsloth	Apache-2.0
sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8/	语音识别 / Speech recognition	csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8	NVIDIA (nvidia/parakeet-tdt-0.6b-v2); ONNX export by the sherpa-onnx project	CC-BY-4.0
nemo_en_titanet_small.onnx	说话人识别 / Speaker labels	csukuangfj/speaker-embedding-models	NVIDIA (TitaNet-small); ONNX export by the sherpa-onnx project	NeMo Toolkit license (Apache-2.0)
silero_vad.onnx	人声检测 / Voice activity detection	csukuangfj/vad	Silero Team	MIT
opus-mt-en-zh-ct2/	草稿翻译 / Draft translation	gaudi/opus-mt-en-zh-ctranslate2	Helsinki-NLP (Jörg Tiedemann et al.); CTranslate2 conversion by gaudi	Apache-2.0

每个模型都按它自己的许可证使用和分发，这个仓库没有为它们设定统一的许可证。Parakeet-TDT（CC-BY-4.0）在使用时需要注明 NVIDIA 为原作者。

Each model is used and redistributed under its own license; this repository does not apply a single license to them. Parakeet-TDT (CC-BY-4.0) requires attribution to NVIDIA.

目录结构 / Layout

和 UnderText 的模型文件夹一致，下载后可以直接使用：

Same layout as UnderText's models folder:

Qwen3.5-2B-Q4_K_M.gguf
sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8/
  encoder.int8.onnx
  decoder.int8.onnx
  joiner.int8.onnx
  tokens.txt
nemo_en_titanet_small.onnx
silero_vad.onnx
opus-mt-en-zh-ct2/
  config.json
  model.bin
  shared_vocabulary.json
  source.spm
  target.spm

引用 / Citations

• Parakeet-TDT: Xu et al., "Efficient Sequence Transduction by Jointly Predicting Tokens and Durations", ICML 2023.
• TitaNet: Koluguri et al., "TitaNet: Neural Model for Speaker Representation with 1D Depth-wise Separable Convolutions and Global Context", ICASSP 2022.
• Opus-MT: Tiedemann and Thottingal, "OPUS-MT — Building open translation services for the World", EAMT 2020.