v3.95
jianchang512/pyvideotransv3.95Jan 25, 2026by jianchang512
AI Summary
Major update with extensive new features including multi-GPU support, multi-process processing, command-line mode, supertonic voice channel, and advanced options for controlling hallucinations. Added document alignment feature and improved speaker diarization.
Key Highlights
- Added secondary recognition in video translation
- Added VAD type switching (ten-vad default, silero alternative)
- Faster-whisper batch inference optimization
- Added supertonic voice channel (EN, KO, ES, PT, FR)
- Voice denoising using Ali modelscope
- Speaker diarization with multiple model support
- Multi-GPU concurrent processing support
- Multi-process for heavy tasks
- Added command-line mode
- Added qwen-tts open source model support
- Added document alignment feature
- Added attach full subtitle option for AI translation
New Features
- Secondary recognition for video translation
- VAD type switching (ten-vad/silero)
- Faster-whisper batch inference
- Supertonic voice channel
- Ali modelscope voice denoising
- Speaker diarization (pyannote3.0, cam++, reverb-diarization)
- Multi-GPU support
- Multi-process processing
- Command-line interface
- Qwen-TTS open source model support
- Document alignment feature
- Full subtitle attachment for AI translation
Full Release Notes
<details> <summary>v3.95更新日志</summary> 1. 在视频翻译功能中,增加二次识别,主界面点击`设置更多参数`可看到 2. 菜单-工具-高级选项中增加`vad类型切换`,默认 `ten-vad`,速度更快更精准,若遇到问题可切换回 `silero` 3. faster-whisper默认批量推理,以加快速度,高级选项中修改`faster-whisper批次`为1,可改为非批量,速度降低但略微提高精确度 4. 优化模型自动下载,先测试`huggingface.co`,若失败则从国内镜像`hf-mirror.com`下载 5. 高级选项中增加几个自定义选项,用于控制幻觉问题 6. 添加 `文稿对齐` 功能,基于 `faster-whisper本地模型`和diff算法简单实现 7. 增加 `supertonic` 配音渠道,支持 英语、韩语、西班牙语、葡萄牙语、法语, 小巧快速 8. 语音降噪改用阿里 modelscope.cn 默认,处理耗时将大幅增加,降噪效果也更佳 9. 说话人分离支持 内置中英模型、pyannote3.0模型、阿里cam++中英模型、reverb-diarization-v1模型,可在高级选项中切换,默认使用内置中英模型 10. 单个视频翻译时,在弹出字幕修改界面时,可打开字幕文件夹,然后手动或通过第三方工具修改已生成的字幕文件,然后在该界面点击`不保存只继续`将使用已做过修改的字幕文件,用于方便自行调整字幕 11. 支持多显卡并发,加快处理速度,可在高级选项中选中`多显卡模式` 12. 重型任务改用多进程方式,以便更好利用多核,可在高级选项中--填写更大数量的`最大进程数`加快速度 13. 高级选项-字幕翻译-增加`附带完整字幕`复选框,选中后并且使用AI翻译渠道时,将把原始字幕一并发给ai,可取得更好的翻译质量 14. 增加命令行模式 [具体使用方法查看cli命令行](https://pyvideotrans.com/cli) 15. 优化默认翻译提示词 16. 增加对 qwen-tts 开源模型支持 </details> ## v3.95 完整包下载(3.95G 含medium模型) > 如果未安装过旧版本,请在此下载完整版,如需cuda加速,需有英伟达显卡并且安装cuda12.8及cudnn9 百度网盘下载地址: https://pan.baidu.com/s/1cEj_LPHGJOZnbcL8xeeRXA?pwd=1234 huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-pyvideotrans-v3.95.7z?download=true ## v3.95 补丁包(380MB) > 不可单独使用,仅用于覆盖旧包,与完整包区别是不含有`torch`模块 > 因torch从 2.5 升级到 2.7,补丁包不含有torch,直接下载覆盖补丁包可能导致无法cuda加速,建议下载完整包覆盖 > **覆盖后如果闪退,请下载完整包** 百度网盘下载地址: https://pan.baidu.com/s/1sgA5rtnn7XGTH_Hl_rjTuQ?pwd=1234 huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-PatchUpdate-v3.95.7z?download=true ## v3.95 Bug修复(sp.exe) v3.95版本bug修复后只更新sp.exe,下载该exe文件后直接覆盖原sp.exe,仅用于3.95版本 https://github.com/jianchang512/stt/releases/download/0.0/sp.exe --- <details> <summary><h3>不喜欢新版?继续下载v3.92旧版本</h3></summary> 百度网盘下载地址: https://pan.baidu.com/s/1iD9ivN8OHaj05nFXNKln8A?pwd=1234 huggingface.co: https://huggingface.co/mortimerme/repocollect/resolve/main/win-pyvideotrans-v3.92.7z?download=true </details>