LectureFlow 3.0.0

好久不见，或者说，你好。 / Long time no see — or rather, hello.
Hello. 你好。哈囉。你好呀。Здравствуйте. नमस्ते.

Although it is numbered 3.0.0, this is LectureFlow’s first public release.
虽然编号是 3.0.0，但这是 LectureFlow 的第一个公开发行版。

This is a pre-release, because macOS and Apple Silicon support is not yet complete. Feel free to try it out, and please open Issues to let me know what you find.
这是一个预发行版，因为 macOS 及 Apple Silicon 设备的支持还不完善。 欢迎各位尝鲜，并且提 Issues 给我看。

You can also support me on GitHub Sponsors so I can buy an Apple Silicon device for testing.
您也可以在 GitHub Sponsors 里支持本人购买 Apple Silicon 设备用于测试。

Baseline devices / 使用的基线设备

• Main machine / 主力机: ASUS ROG Zephyrus G16
  • Intel Core Ultra 9 386H (Xe3 integrated graphics + Intel AI Boost NPU / Xe3 核显 + Intel AI Boost NPU)
  • NVIDIA GeForce RTX 5070 Ti Laptop GPU
  • 64 GB RAM, Windows 11 / 64 GB 内存，Windows 11
• Remote GPU / 远端 GPU: Google Colab · T4 (paid compute / 付费额度)
• macOS: passed automated tests on GitHub Actions (Apple Silicon); not yet verified on a real Mac.
  仅在 GitHub Actions 的 Apple Silicon 环境中通过自动化测试，尚未在真机上验证。

What it does / 目前实现的功能

From recording to transcript / 从录像到讲稿

• Transcribes lectures on your own computer, sentence by sentence, with timestamps.
  在本机转写讲座录音，生成逐句、带时间戳的文稿。
• Builds a course glossary from your syllabus and slides, cleaned up by a second “janitor” model, so technical terms are right from the first sentence.
  根据课程大纲和课件整理术语表，再由另一个模型“清道夫”清理，专业术语从第一句起就是对的。
• Corrects misheard words and translates sentence by sentence, using a summary of the whole lecture and the surrounding sentences. Changes it isn’t sure about are flagged so you can check the original audio.
  结合整节课的摘要和前后文修正听错的地方、逐句翻译；没把握的修改会标出来，可以点开听原声确认。
• Writes a bilingual title, a one-sentence summary, and chapters for each lecture.
  自动生成每节课的中英文标题、一句话简介和章节。

A bilingual player built for studying / 为学习设计的双语播放器

• Subtitles sit below the video and never cover the slides or the board.
  字幕放在视频下方，不遮挡课件和板书。
• The transcript scrolls with playback; click any sentence to jump to it. Previous / next sentence, loop a sentence, slow down.
  逐句文稿随播放滚动，点击即跳转；上一句、下一句、单句循环、慢速播放。
• Picks up where you left off; copy a citation accurate to the second.
  关掉再打开，从上次的位置继续；一键复制精确到秒的引用。

Course organisation / 课程整理

• An AI-written rule reads recording dates from your file names; lectures are numbered by calendar (e.g. W02L1: week 2, session 1) and grouped by teaching week.
  由 AI 生成文件名解析规则，按录制日期自动编号（如 W02L1：第 2 周第 1 节），按教学周分组。
• Tells lectures, tutorials, and seminars apart.
  区分讲座、辅导、研讨等课程类型。

Devices and power awareness / 转写设备与电源感知

• Supports NVIDIA GPUs (CUDA), Intel integrated graphics and NPUs (OpenVINO), and Apple Silicon (MLX, experimental).
  支持 NVIDIA 显卡（CUDA）、Intel 核显与 NPU（OpenVINO）、Apple Silicon（MLX，实验性）。
• Laptops use one transcription device at a time: the dedicated GPU when plugged in, the NPU on battery. Tasks that can’t move are paused until you plug back in.
  笔记本一次只用一台转写设备：插电用独显，拔电自动换到 NPU；无法迁移的任务会挂起，接回电源后继续。
• Optionally borrow a Google Colab GPU: LectureFlow starts it, transcribes, downloads the results, and shuts it down, within a 4-hour fair-use window.
  可选借用 Google Colab 的 GPU：自动开机、转写、下载结果、关机，遵守 4 小时的公平占用窗口。
• Device self-test: measures speed with a clip of your own lecture and recommends a model and precision.
  设备自测：用你自己的一段讲座测速，推荐合适的模型和精度。

Desktop app / 桌面应用

• Installers for Windows and macOS with a guided first-run setup; no terminal needed. Transcription components are downloaded only for the hardware you have.
  Windows 与 macOS 安装包，首次启动有引导，不需要命令行；转写组件按硬件按需下载。
• Keeps your computer awake while it’s working.
  处理期间自动阻止电脑休眠。
• Interface in English, 简体中文, 繁體中文（台灣）, 粵語（香港）, Русский, हिन्दी.
  界面支持 English、简体中文、繁體中文（台灣）、粵語（香港）、Русский、हिन्दी。

Export / 导出

• Bilingual transcripts, bilingual subtitles (VTT), and a JSON file with timestamped citations for other AI tools to work with.
  中英对照讲稿、双语字幕（VTT），以及带时间戳引用的 JSON，方便交给其他 AI 整理笔记。

Known limitations / 目前的局限

• The setup guide has some minor known issues.
  用户引导界面有轻微的已知问题。
• Scheduling across multiple inference backends is not yet smart.
  多推理后端的调度还不够智能。
• AMD ROCm is not supported yet.
  AMD ROCm 尚未得到支持。
• macOS and Apple Silicon have not been verified on a real Mac.
  macOS 与 Apple Silicon 尚未在真机上验证。
• Installers are not code-signed; Windows and macOS will show security warnings. See docs/install.md for how to continue.
  安装包未经代码签名，Windows 与 macOS 会弹出安全提示，放行方法见 docs/install.md。
• This release does not update itself; download new versions manually.
  本版本不会自动更新，新版本发布后需要手动下载。
• So far tested only by computer science students, with English lectures translated into Simplified Chinese.
  目前只由计算机专业的学生测试过，讲座语言为英语、翻译目标为简体中文。