Skip to content

LectureFlow 3.1.0

Choose a tag to compare

@github-actions github-actions released this 26 Oct 19:41
· 4 commits to main since this release

LectureFlow 3.1.0 "Greggs" has arrived!

After renting a painfully awkward Mac mini, Claude helped me fix things one by one.
This is LectureFlow's first official release.

New features

  • Brand-new first-run setup: choose how to transcribe first (on this device, or on a remote Google Colab GPU), then choose where to keep your data (an external drive works). Once you confirm the location, components start downloading in the background right away. (The data directory can now be changed in the wizard, thanks to @Wineondili.)
  • One-step Colab setup: install the Google Cloud CLI and sign in to Google from inside the app, no terminal needed.
  • The "Transcription only" tier now really works: transcribe, play and download English subtitles without configuring any model provider.
  • AV1 videos can be transcoded for playback: the bundled ffmpeg now includes dav1d.

Improvements

  • MLX: devices that support MLX now correctly use MLX, and only MLX.
  • macOS support: the Diagnostics page checks MLX the Mac way; the resource bar shows CPU, unified memory and GPU; battery and Low Power Mode are supported; CPU devices that can't actually be used are no longer shown.

Fixes

  • Power source not detected correctly on macOS (thanks @Mashirl)
  • Power awareness not reported correctly on macOS (thanks @Mashirl, and an anonymous friend who lent a MacBook Air)
  • The Diagnostics page didn't handle backends that don't exist

Known issues

  • The Whisper large-v3 family (including turbo, the Mac default) often drops punctuation and capitalisation. This is a known issue with experiments scheduled.
  • Not yet tested on a MacBook running on battery. Please open the Sponsor page and help me buy one.

LectureFlow 3.1.0 "Greggs" 参上!

在租了一台非常难用的 Mac mini 之后,感谢 Claude 大小姐帮忙一项项修好了。
这是 LectureFlow 的第一个正式 Release。

新功能

  • 全新首次引导:先选转写方式(本机转写或 Google Colab 远端 GPU),再选数据存放位置(可以放到外置盘)。确认位置后,组件会立即在后台下载。(数据目录可在向导中修改,感谢 @Wineondili)
  • Colab 配置一步到位:在应用内一键安装 Google Cloud CLI 并登录 Google,不用再自己开终端。
  • “只转写”档位真正可用:不配置任何模型服务,也能转写、播放和下载英文字幕。
  • AV1 视频可以转码播放:内置的 ffmpeg 加入了 dav1d。

改进

  • MLX:支持 MLX 的设备现在会正确地使用、且只使用 MLX。
  • macOS 适配:诊断页按 Mac 的方式检查 MLX;资源栏显示 CPU、统一内存和 GPU;支持电池与低电量模式;不再显示实际用不了的 CPU 设备。

修复

  • macOS 未能正确检测电源(感谢 @Mashirl)
  • macOS 未能正确表达电源感知(感谢 @Mashirl,以及一位不具名朋友出借的 MacBook Air)
  • 诊断页无法正确处理不存在的后端

已知问题

  • Whisper large-v3 系列(包括 Mac 上默认的 turbo)经常漏掉标点和大小写,这是一个在排期试验的已知问题。
  • 还没有在使用电池供电的 MacBook 上实测。请点开 Sponsor 页面支持我买一台。