Skip to content

AISRT v0.1.3

Choose a tag to compare

View all tags
@zkwi zkwi released this 05 May 23:10
· 3 commits to main since this release
v0.1.3
37d1474

AISRT v0.1.3

简体中文

本次发布提供轻量 Windows portable 源码包、Python wheel，以及实验性的 Windows x64 CUDA 完整运行时安装包。主要变化仍是优化本地 ASR 默认参数，在不切换模型、不提高批大小的前提下改善常规处理速度。

主要变化

  • 将 ASR 每个音频块的默认 max_new_tokens 从 2048 降到 1536，减少异常长生成带来的耗时。
  • 保留稳定的默认组合：1.7B ASR 模型、45 秒音频分块和 batch size 1。
  • CLI 帮助、GUI 推荐模式和底层模型加载默认值已统一使用新的生成上限。
  • 中英文 README 增加 --max-new-tokens 说明，长对白被截断时仍可临时调高。
  • 增加 CLI 和 GUI 默认参数测试，降低后续改动误改推荐档位的风险。

发布资产

  • AiSRT-v0.1.3-windows-x64-cuda-runtime-setup.exe：实验性 Windows x64 CUDA 完整运行时安装包，包含 Python、PyQt6、PyTorch/CUDA 运行库和 AISRT 程序，不包含模型权重和 FFmpeg。
  • AiSRT-v0.1.3-windows-portable.zip：轻量 Windows 包，包含源码、启动脚本和依赖文件。
  • aisrt-0.1.3-py3-none-any.whl：Python wheel，适合手动安装的用户。
  • SHA256SUMS.txt：发布资产 SHA256 校验值。

注意事项

  • 完整运行时安装包不包含模型权重、模型缓存、FFmpeg、媒体文件、生成字幕、截图、测试文件或日志。
  • 完整运行时安装包仍需要 NVIDIA CUDA 显卡和兼容的 NVIDIA 驱动；用户不需要单独安装 CUDA Toolkit。
  • 使用远程模型 ID 时，AISRT 会在首次运行时下载模型权重到配置的 Hugging Face 缓存目录。
  • portable ZIP 不包含 Python、PyTorch/CUDA 运行库、模型权重、模型缓存、FFmpeg、媒体文件、生成字幕、截图、测试文件或日志。
  • 使用 portable ZIP 时，解压后运行 install_runtime.bat，脚本会创建 .venv 并安装 Python 依赖。
  • FFmpeg 和 ffprobe 需要单独安装，并确保可在 PATH 中直接运行。

English

This release provides a lightweight Windows portable source package, a Python wheel, and an experimental Windows x64 CUDA full-runtime installer. The main change remains tuning local ASR defaults for better everyday speed without changing the model or raising batch size.

Highlights

  • Lowered the default ASR max_new_tokens per audio chunk from 2048 to 1536 to reduce time spent on unusually long generations.
  • Kept the stable default combination: 1.7B ASR model, 45-second audio chunks and batch size 1.
  • Unified the CLI help, GUI recommended profile and model-loading default around the new generation limit.
  • Documented --max-new-tokens in both README files; users can still raise it temporarily if long dialogue is truncated.
  • Added CLI and GUI default-parameter tests to keep the recommended profile from drifting accidentally.

Assets

  • AiSRT-v0.1.3-windows-x64-cuda-runtime-setup.exe: experimental Windows x64 CUDA full-runtime installer with Python, PyQt6, PyTorch/CUDA runtime libraries and AISRT, excluding model weights and FFmpeg.
  • AiSRT-v0.1.3-windows-portable.zip: lightweight Windows package with source code, launch scripts and dependency files.
  • aisrt-0.1.3-py3-none-any.whl: Python wheel for users who prefer manual Python installation.
  • SHA256SUMS.txt: SHA256 checksums for release assets.

Notes

  • The full-runtime installer does not include model weights, model caches, FFmpeg, media files, generated subtitles, screenshots, tests or logs.
  • The full-runtime installer still requires an NVIDIA CUDA GPU and a compatible NVIDIA driver; users do not need to install the CUDA Toolkit separately.
  • When remote model IDs are used, AISRT downloads model weights to the configured Hugging Face cache on first use.
  • The portable ZIP does not include Python, PyTorch/CUDA runtime libraries, model weights, model caches, FFmpeg, media files, generated subtitles, screenshots, tests or logs.
  • For the portable ZIP, run install_runtime.bat after extracting it to create .venv and install Python dependencies.
  • FFmpeg and ffprobe must be installed separately and available in PATH.
Assets 6
Loading