v4.12
microsoft/TinyTroupev4.12Sep 6, 2026by jianchang512
Upgrading?
AI Summary
This release introduces new ASR models and removes legacy features from the Whisper pipeline to improve language support and stability.
Key Highlights
- Added `VibeVoice-ASR` voice recognition channel
- Replaced `parakeet日语` with `Nemotron-3.5-asr` supporting 40 languages
- Added `nvidia/parakeet-tdt-0.6b-v3` model for European languages
- Removed `whisper预分割音频` and `合并邻近短字幕` features
- Optimized VAD (Voice Activity Detection) algorithm
Breaking Changes
- Removed `whisper预分割音频` (Whisper pre-split audio)
- Removed `合并邻近短字幕` (Merge adjacent short subtitles)
New Features
- Upgrade `transformers` to 5.13.1
- New `LLM纠错` (LLM Correction) feature using DeepSeek
- Fixed Qwen-asr language setting errors
Full Release Notes
## v4.12 Win10/11 预打包版本下载 <details> <summary>MacOS/Linux系统请使用源码部署</summary> 1. 安装依赖: ```bash # macOS brew install libsndfile git brew tap homebrew-ffmpeg/ffmpeg brew install homebrew-ffmpeg/ffmpeg/ffmpeg # Ubuntu/Debian sudo apt-get install ffmpeg libsndfile1-dev ``` 3. 安装 uv: ```bash curl -LsSf https://astral.sh/uv/install.sh | sh ``` 4. 克隆并启动: ```bash git clone https://github.com/jianchang512/pyvideotrans.git cd pyvideotrans uv sync uv run sp.py ``` </details> ### 更新日志/Changelos - 修复 Qwen-asr(内置) 语言设置报错问题 - 增加 `VibeVoice-ASR` 语音识别渠道 - `parakeet日语(内置)`识别渠道替换为`Nemotron-3.5-asr-0.6b(内置)`模型,支持欧洲系及日韩越等40种语言 - `Huggingface_ASR(内置)`增加`nvidia/parakeet-tdt-0.6b-v3`模型,支持10多种欧洲语言 - 移除`whisper预分割音频`、`合并邻近短字幕` - 优化 VAD 切分算法 - 优化翻译提示词 - `LLM重新断句`调整为`LLM纠错`,只修改错别字和恢复标点,不重新断句,可在`高级选项-通用设置-LLM纠错AI渠道`调整,默认 DeepSeek - `二次识别`固定使用`faster-whisper`,以增强语言支持和兼容性,默认使用`large-v3-turbo`模型,可在`高级选项-语音识别参数区域-二次识别所用模型`调整 - `语音转录/文字合成/翻译字幕` 独立功能显示全部可用的80种语言 - 视频翻译发音语言下拉列表增加`自动检测` - GeminiAI语音识别渠道固定使用`gemini-3.5-transcribe`模型 - 升级 `transformers`到 5.13.1 - 修复其他已知bug ---- ### v4.12 完整包下载(2.7G) - 百度网盘地址: https://pan.baidu.com/s/11LKtD3l5Ti-H2DprXRqNrg?pwd=1234 - huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-pyvideotrans-v4.12-0913.7z?download=true ### v4.12 补丁包下载(360MB) > 补丁包不可单独使用,仅用于覆盖旧包, 如果版本相差较多,请直接下载完整包,勿使用补丁包。覆盖后如果闪退或报错,请下载完整包。 - 百度网盘地址: https://pan.baidu.com/s/19vRD9nytjLWudu1zhOFNhA?pwd=1234 - huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-PatchUpdate-v4.12-0913.7z?download=true ---- <details><summary>解压注意事项</summary> - 解压7z压缩包,解压后,找到 `sp.exe` 文件,双击即可启动软件,不可在压缩包内双击运行。 - 不要将软件解压到桌面或 C 盘的 `Program Files` 等需要管理员权限的文件夹下。 - 避免路径错误: 解压路径中不要包含中文、空格或特殊符号。 - 如果使用7zip软件解压提示错误,请尝试使用 [Bandizip](https://bandisoft.com/bandizip/) 或 [360压缩软件](https://yasuo.360.cn) 进行解压 </details> ---- <details><summary> 如果新版遇到问题可继续使用上个版本(v4.11/v4.03)</summary> **v4.11 完整包下载(2.7G)** - 百度网盘地址: https://pan.baidu.com/s/1GkL4pyAYxJRvRor0jfh2rg?pwd=1234 - huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-pyvideotrans-v4.11.7z?download=true **v4.09 完整包下载(2.7G)** - 百度网盘地址: https://pan.baidu.com/s/1deSACf5Wmy34Ua8hzgDQVw?pwd=1234 - huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-pyvideotrans-v4.09-0816.7z?download=true **v4.03 完整包下载(2.7G)** - 百度网盘地址: https://pan.baidu.com/s/1yENT4t7E5K1OrbOFS7WEvA?pwd=1234 - huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-pyvideotrans-v4.03-0626.7z?download=true </details> #### New Contributors * @Biga73 made their first contribution in https://github.com/jianchang512/pyvideotrans/pull/1194 **Full Changelog**: https://github.com/jianchang512/pyvideotrans/compare/v4.11...v4.12
Related Repositories
rasbt/LLMs-from-scratch★ 107,043
Implement a ChatGPT-like LLM in PyTorch from scratch, step by step
lyogavin/airllm★ 40,152
AirLLM 2.8T LLM inference with single 4GB GPU
NirDiamant/RAG_Techniques★ 29,958
This repository showcases various advanced techniques for Retrieval-Augmented Generation (RAG) systems. Each technique has a detailed notebook tutorial.
microsoft/OmniParser★ 25,539
A simple screen parsing tool towards pure vision based GUI agent
karpathy/nn-zero-to-hero★ 24,906
Neural Networks: Zero to Hero