v0.3.0-alpha.4
Pre-release
课溯 · VeriLecture v0.3.0-alpha.4
把课堂录音变成可核对的复习重点。
Trace every key point back to the lecture.
This is the second public V3 milestone. It publishes the desktop shell for three platforms while keeping native local-ASR support and the CUDA Runtime status explicit.
这次发布能做什么
- 导入课堂录音，并保留原始音频和原始转写。
- 使用课程词库进行校准；修订和编辑生成带来源的新版本。
- 在文本整理服务获得授权后生成复习重点，并保留相关转写片段与时间点。
- 从重点或转写位置回到原音，或导出 Markdown、JSON 和纯文本结果。
- 在首次运行和设置中查看硬件、本地转写和数据授权边界。
What this release makes possible
- Import classroom audio while keeping the original recording and raw transcript available.
- Calibrate terminology with a course lexicon; corrections and edits create new versions with provenance.
- Generate review points through a consented text service and retain the related transcript segments and timestamps.
- Return from a point or transcript line to the source position, then export Markdown, JSON, or plain text.
- See hardware, local transcription, and consent boundaries during first run and in Settings.
Package and runtime status
| Platform | Package | Desktop package | Native local ASR |
|---|---|---|---|
| Windows x64 | NSIS installer | Published | Fun-ASR route and a representative CUDA path have corresponding validation |
| Linux x64 | AppImage | Published | Native sidecar pending publication and validation |
| macOS | DMG | Published | Native sidecar pending publication and validation |
The existence of a desktop package does not prove that local ASR is ready on that platform.
Documentation note
The public README and website include a generated result-screen concept mockup marked 概念示意 · 非实机截图. It explains the intended relationship between points, transcript segments, timestamps, and playback; it is not evidence of a verified Windows UI or runtime path. A real result capture remains pending the acceptance path in the screenshot brief.
Important limitations
- This is an Alpha release. Back up important recordings before installing.
- Linux and macOS packages provide the desktop application, but native local-ASR sidecars still require separate publication and validation.
- The Qwen3-ASR tiers require a compatible NVIDIA CUDA path and a separately hosted CUDA Runtime. The approximately 4.6 GB Runtime is excluded from Git and this Release; the installer keeps the download gated until a stable HTTPS artifact is published and accepted.
- Model weights are downloaded on first use and require a network connection, free disk space, and an integrity check.
- Long audio is processed as a bounded local job; durable in-job resume is not exposed in the current interface.
- Text-model features are optional, require a configured provider and explicit consent, and have narrower provider and long-document coverage while the Alpha is being expanded.
- The Windows installer is unsigned unless signing is supplied by the release environment.
Read the full list in Known Limitations.
Checksums
Each installer has a matching SHA256 checksum file. The checksum entry uses the same filename as the published asset. Verify the matching file before installing:
First run
- Read the privacy boundary and choose the consent scope.
- Let VeriLecture scan hardware, CUDA, storage, and the model directory.
- Choose a local transcription route that the device can run.
- Test a short recording before importing a full lecture.
Choose Fun-ASR-Nano-2512 when no usable NVIDIA GPU is available. For Qwen, wait for a Release that explicitly says the CUDA Runtime is publicly hosted and accepted.
License and attribution
VeriLecture is released under the MIT License. Third-party licenses and upstream attribution are listed in NOTICE and THIRD_PARTY_NOTICES.md.