LectureRelay 0.3.13 — 处理方式回退修正

下载 LectureRelay_0.3.13_x64-setup.exe，安装后即可使用，无需自行构建。运行安装包即可从旧版升级，课程、课堂记录和已下载模型会保留。此版本仍为未签名的 Windows x64 开发预览版。使用 0.3.12 的用户建议升级。

本次更新

• 修正 0.3.12 的一处设置读取问题：已保存 Groq 或 OpenAI 服务时，若偏好设置中“课堂翻译”或“课后学习”的处理方式是当前版本无法识别的值，0.3.12 会把它误当作旧版本缺少的字段，改为云端处理；配置了 API Key 时，后续文字可能被上传。现在无法识别的处理方式一律回退为本地，只有旧版本确实没有该字段时，才沿用原先明确选择的云端服务。
• 新增回归测试，覆盖无法识别的值、空值和类型错误。
• 开发文档更正：安装包验证脚本会安装到日常使用的安装目录，并非隔离的临时目录。

数据库结构、模型版本和第三方依赖均未改变。其余功能与 0.3.12 相同，见 0.3.12 更新说明。

发布验证

• 安装包源码：44cfd77fad33f736b59faa73c1b12731b5158e7f。v0.3.13 指向此提交；后续文档提交不改变安装包。
• Windows checks：八组 React 回归、88 个桌面 Rust 测试、1 个环境检查测试、类型检查、格式、Clippy 与仓库检查。新增的回归测试在恢复 0.3.12 逻辑时失败（得到 Cloud 而非 Local），修正后通过。
• 干净 Windows 安装包构建：全新 GitHub Windows runner 编译，完成 NSIS 安装、运行资源、安装内容、图标和许可校验。安装包 16,321,228 bytes，SHA-256 c68fa389646f7de934f930862b0fc306383055ff5d3f48f6930e48563c713910。
• 本次未播放音频、发起真实云端 API 请求、重跑安装版完整课堂流程或 90 分钟测试。

查看 简体中文说明、首次使用指南 和 验证说明。旧版 Release 和安装包保持原内容。

English

0.3.13 fixes a 0.3.12 settings issue: when a Groq or OpenAI service was saved and the stored translation or postclass study mode was a value this version does not recognize, 0.3.12 treated it as a missing legacy field and switched it to cloud processing, which could upload text if an API key was configured. Unrecognized modes now fall back to Local; only a field that is genuinely absent from older preferences keeps a previous explicit cloud selection. A regression test covers unknown, null and wrong-type values. The EXE is built from the source commit above on a clean GitHub Windows runner. Install the EXE; no developer tools or LectureRelay account are required.

LectureRelay 0.3.12 — 提示准确性与内部结构整理

下载 LectureRelay_0.3.12_x64-setup.exe，安装后即可使用，无需自行构建。运行安装包即可从旧版升级，课程、课堂记录和已下载模型会保留。此版本仍为未签名的 Windows x64 开发预览版。

本次更新

• 一批原先只显示“操作未完成，请重试”的提示，现在显示准确的中文说明，涉及实时总结读写、服务商错误和退出时中断的任务等。
• 指向已移除菜单“设置 → 本地 AI”“AI 服务商”的提示，已更正为“设置 → AI 与模型 → 本地模型”等当前位置。
• 导入音视频与其他 AI 任务一致，会等待模型下载完成；云端实时总结不使用本地模型，不再阻止下载或删除模型。
• 偏好设置中出现无法识别的值时回退为默认值，不再导致设置读取失败；无法识别的实时总结服务会被关闭，不会改为云端服务。
• 内部整理：实时字幕拆分为识别与翻译两个模块，互斥判断集中到一处，设置与状态字段类型化，界面文字全部进入中英文目录，架构文档按当前代码重写。

数据库结构、模型版本和第三方依赖均未改变。本地 Qwen 实时总结仍停用；课后手动复习与问答仍可使用已下载的 Qwen。

发布验证

• 安装包源码：eb7c54d4bd6ad2f8ad3d0544514626c82a32de80。v0.3.12 指向此提交；后续文档提交不改变安装包。
• Windows checks：八组 React 回归、87 个桌面 Rust 测试、1 个环境检查测试、类型检查、格式、Clippy 与仓库检查（含界面文案与原生提示翻译检查）。
• 干净 Windows 安装包构建：全新 GitHub Windows runner 编译，完成 NSIS 安装、运行资源、安装内容、图标和许可校验。安装包 16,322,292 bytes，SHA-256 b6d0f0a19e54ccb72f1337b43aab66d3d3adf4b0b73603526acf9c7680eec04f。
• 真实音频直播会话测试（开发机）：将 MIT OCW 9.00 第 1 讲 90 秒真人讲课片段按实时速度写入录音，经完整直播会话运行本地 Nemotron 识别与 Hy-MT2 翻译。得到 7 段定稿、238 个英文词，7 段全部完成翻译，识别跟上实时进度；英文结果与重构前版本逐字节一致。此测试不经过声卡采集，也不是安装版完整课堂流程或 90 分钟测试。

查看 简体中文说明、首次使用指南 和 验证说明。旧版 Release 和安装包保持原内容。

English

0.3.12 shows accurate Chinese messages where several errors previously fell back to a generic notice, fixes hints that pointed to removed settings sections, makes media import wait for model downloads, no longer blocks model management during cloud summaries, and tolerates unknown stored preferences. Internally, live captions are split into recognition and translation modules, busy checks are centralized, settings are typed, and all interface text lives in checked catalogs. A real 90-second lecture run through the full live session (local Nemotron and Hy-MT2) produced English byte-identical to the previous code. The EXE is built from the source commit above on a clean GitHub Windows runner. Install the EXE; no developer tools or LectureRelay account are required.

LectureRelay 0.3.11 — 设置整理与状态可靠性

下载 LectureRelay_0.3.11_x64-setup.exe，安装后即可使用，无需自行构建。运行安装包即可从旧版升级，课程、课堂记录和已下载模型会保留。此版本仍为未签名的 Windows x64 开发预览版。

本次更新

• 设置整理为六个主分类，首次设置独立保留。声音设备和字幕样式集中管理；AI 功能、本地模型、实时总结位于同一分类，API Key 和高级选项默认折叠。
• 外观和安静模式自动保存；其他偏好有改动时才出现保存栏。字幕、课程与学习功能的设置按钮直接跳到相关选项。
• 按职责拆分设置组件，使用稳定导航标识和有类型的草稿合并；独立保存运行时或总结偏好时，保留其他尚未保存的修改。
• 只读取当前声音来源的设备；一种设备读取失败不会阻塞另一种，迟到响应不会覆盖新选择，连点声音测试只发起一次预览。
• 清理存储的主要错误与刷新失败分别显示，保留永久删除和清空数据的输入确认。

本地英文识别和翻译、云端实时总结、手写笔记继续使用原有流程。本地 Qwen 实时总结仍停用；课后手动复习与问答仍可使用已下载的 Qwen。第三方依赖、模型版本与已发布 SQL 迁移均未更新。

发布验证

• 安装包源码：4575dc5894281c70048205afbd569abd8fea2152。v0.3.11 指向此提交；后续文档提交不改变安装包。
• Windows checks：八组 React 回归、81 个桌面 Rust 测试、1 个环境检查测试、类型检查、格式、Clippy 与仓库检查。
• 干净 Windows 安装包构建：全新 GitHub Windows runner 编译，完成 NSIS 安装、运行资源、安装内容、图标和许可校验。附件同时提供 SHA-256 与 build-metadata.json，可核对源码和构建来源。
• 本轮没有播放声音、发起真实 API 请求、重新进行安装版完整课堂流程或 90 分钟测试。React 测试使用隔离响应，不能代替真实音频与模型推理验收。

查看 简体中文说明、首次使用指南 和 设置与工程审查。旧版 Release 和安装包保持原内容。

English

0.3.11 groups settings into six categories, links each feature to the relevant options, preserves unsaved preferences during independent saves, isolates audio-device failures and stale responses, suppresses duplicate input tests, and keeps storage operation errors separate from refresh failures. The EXE is built from the source commit above on a clean GitHub Windows runner. UI regressions and installer verification are separate from real classroom acceptance. Install the EXE; no developer tools or LectureRelay account are required.

LectureRelay 0.3.10 — 实时课堂要点与界面精简

LectureRelay 0.3.10 — 实时课堂要点

下载 LectureRelay_0.3.10_x64-setup.exe，安装后即可使用；已有安装版可直接升级。无需自行构建，也不要求注册 LectureRelay 账户。本版本为未签名的 Windows x64 开发预览版。

本次相对于 0.3.8 的主要变化：

• 新增右侧实时分段总结卡片：默认关闭，启用后每四分钟整理新定稿英文，也可选择两分钟、五分钟或手动整理。
• Groq 为优先提供的可选云端入口，初始模型为 openai/gpt-oss-120b；OpenAI 也可用。获取 Key、保存、测试连接、确认上传、启用组成完整配置流程。总结设置不改变本地英文识别和翻译。
• 卡片保留时间范围、原文与引用回听入口；手写笔记独立保存，仅在用户点击时追加 AI 要点。停止录音先保存音频，不等待总结，失败或未完成内容保留明确状态。
• 停用本地 Qwen 实时总结及其原生推理路径。旧本地总结设置自动关闭、清除上传同意，不会自动切到云端。已有卡片、模型文件、转录和笔记保留；课后手动复习和问答仍支持 Qwen。
• 修复云端客户端在全新进程中的 TLS 初始化，以及 JSON 浮点时间戳往返导致未改动的总结被误判过期的问题。
• 设置入口换成齿轮，清理课程、字幕、总结、时间索引和笔记里的常驻解释文案；仅在确有待翻译内容时显示数量，保留实际状态、错误与确认提示。

使用 Groq 实时总结

进入 设置 → 实时总结，依次完成 获取 Groq API Key → 粘贴并保存 → 测试总结连接 → 确认文字上传 → 启用实时总结。API Key 由 Windows Credential Manager 保存。

此功能发送所选英文片段与有限的课程背景、术语，不上传录音。Groq 免费账户受请求和 tokens 限额约束；OpenAI API 单独计费。费用与额度由用户自己的服务商账户承担。

构建与验证

• 发布源码：5a00ac5be6c19b420d3f4fde802cfbb7e2dadd8e。
• Windows checks：类型、构建、React、常规 Rust、Clippy、格式与仓库检查。
• 干净 Windows 打包：不恢复项目或依赖缓存；构建原生 worker、应用与 NSIS，实际安装并核对资源、图标、许可和 payload。
• 附件 build-metadata.json 记录此次 CI 产物的源码、工具链、输入及 payload 哈希、安装包大小和验证范围。发布的是该 CI 产物，与此前本地编译的 EXE 不混用。
• 此前 0.3.10 实机设置验收使用精简界面前的本地编译安装包。最终界面另通过中文与实时总结 React 组件测试；本次 CI 附件没有重新执行真实课堂音频、模型推理、Groq 请求或 90 分钟验收。

AI 转录、翻译与总结仍可能出现遗漏或含义错误；引用有效不代表结论一定正确。请保留原文与录音作为核对依据。普通笔记本功耗与风扇噪音尚未完成验收。

升级保留课堂资料与已下载模型。此前 Release 和附件不覆盖。详细说明：简体中文 README、实时总结指南、此前的实机验收记录。

LectureRelay 0.3.8 — Local caption preview

LectureRelay 0.3.8 preview

本版本保留实际 0.3.7 的中文界面、清晰的模型下载状态、可关闭重开的字幕窗、单课堂删除和可恢复复习，继续优化本地 Nemotron + Hy-MT2 的课堂字幕。

• 本地识别更新间隔从 640 ms 缩到 320 ms，500 ms 句尾静音按连续采样累计，避免更新变快后提前拆句。
• 稳定的临时翻译前缀从 32 词扩到最多 80 词；长句后半可以继续更新。修订身份、过期结果拒绝、最终任务优先和有界队列保留。
• 实时翻译使用 4,096 token 上下文和更短提示；课后任务保留 8,192。全零输入且未进入语音段时跳过识别编码，录音与非零低音量仍保留。
• 可选诊断记录准备、执行、修订和真实 UI 出现时间，避免混用首个片段和完整含义的延迟。

同一五分钟 MIT 自然课堂、安静模式的安装版暖启动对照：首英文约 2.17 → 2.33 秒，中文课程主题约 9.10 → 6.68 秒，完整对比含义约 9.33 → 8.97 秒；最大英文积压 2.07 → 1.87 秒。首英文没有明显提速，不能用不同原文长度的首条最终译文宣称大幅改善。播放前静音 CPU 明显减少，讲话期间 CPU 仅小幅降低。

真实麦克风测试有明显漏词，首英文约 9 秒；翻译仍有术语、主语和数字维度错误。日／韩检查仅验证功能连接。最终同素材完整冷启动复测未完成；用户要求先不做新的 90 分钟验收。此包是短测预览，不是完整课堂准确性、长时或普通笔记本验收通过的版本。

下载 EXE 安装即可，无 LectureRelay 账号、Python/Conda/CUDA 或 GPU 要求。本地模型在应用内另行下载；不自动回退云端，不包含实验 Mozilla 权重。安装包未签名。升级保留课堂资料，已有 Release 资产不覆盖。

实际测试安装包：LectureRelay_0.3.8_x64-setup.exe，16,224,824 字节，SHA-256：

504e940b2190e556c0ace7c34fac1ae1fe4bf53a06325b8bd9e12ba79b9155a4

安装后的 EXE SHA-256：c143314ed33d85c5277d9f0f666bc80823140811b7e7aaee252eacecb2517e81。测试的是本地实际安装包；同产品源码的 GitHub Windows cold packaging 另行通过，不声称两份产物逐字节相同。

完整出处、阶段配对、资源、质量错误和验收范围见 0.3.8 研究与安装版报告。

发布标签对应已合并的 main：c59abe8f216e3a1e4ad8c40052af2de483c884f8；产品树与通过检查的 1e466ff 相同。普通 Windows 检查：https://github.com/Eren-Tian/LectureRelay/actions/runs/37392865587 。冷打包：https://github.com/Eren-Tian/LectureRelay/actions/runs/37392864558 。附件 build-metadata.json 对应本地实测安装包，不混用 CI 的另一份二进制。

LectureRelay 0.3.3 - Audio continuity preview

Install on Windows

Download LectureRelay_0.3.3_x64-setup.exe from Assets, run the installer and open LectureRelay. You do not need to build the project or install Node.js, Rust or other developer tools. Existing library data and preferences live separately from the installed program and are preserved on upgrade.

This remains an unsigned Windows x64 preview, with no mandatory LectureRelay account. Download local models once in Settings. This repository is private, so GitHub downloads require repository access. Optional cloud services use your own provider and API key.

Changes since 0.3.2

• Protect recordings during brief writer stalls. The old queue held only 640 ms on the tested audio device; real use reproduced 1.29 seconds of missing audio. Capture now has a bounded five-second mono-sample budget plus a packet limit. Enqueue remains nonblocking, and any overflow remains explicitly counted and visible.
• Retain useful audio evidence. Saved diagnostics distinguish device discontinuity flags, exact application discards, lifecycle transitions and slow writer operations. A device flag does not invent an exact missing-sample count. Warnings remain visible in replay.
• Automated Windows checks. Every main push/PR runs frontend types/build, real React regressions, ordinary Rust tests, Clippy, formatters and repository checks.
• Cold Windows installer builds. Pinned tools and hash-verified native runtimes build and verify the installer on a fresh GitHub runner without restored dependency/project caches or model weights. Release assets now include build and installed-payload metadata.

Validation and limits

The two-second writer-stall regression lost 65,280 samples before the fix and retained all 96,000 samples afterward. Ordinary CI passed 53 application Rust tests plus one environment test, three real React fixtures and Clippy with warnings denied. Hardware/model/credential opt-ins remain separate.

The installed acceptance candidate from 59503ca754ab6c9d1cae3900d28e52b0d3fafd21 completed 90m05.11s with real Windows playback/capture, local speech, local translation and Quiet Mode. No WebDriver/native IPC polling, forced GC, worker restart or state reset was used during capture. All 21,600 ordinary and 3,440 event-neighborhood probes matched at a constant offset; no application audio discard was counted. One startup device flag was retained. The raw WAV is 91m40.84s, including captured lead/tail while the controller completed Stop. All 271 bilingual segments, five timestamp playback checks, restart and original-data/preference preservation passed.

The published rebuild uses the same product/build inputs; subsequent source changes affect only tests and documentation. The published EXE has not separately repeated the 90-minute run, and binary identity with the acceptance candidate is not claimed. The exact tested candidate installer SHA-256 is A29AB6368B9EE2D3FBCE69C6155D75E963CE99B386BF95A06F457063B2FFBB65.

Probe comparisons have finite resolution and cannot certify every sample or perceptual quality. Device flags still cannot reveal exact upstream loss. The internal cause of slow status emissions remains uncertain; a stall longer than the buffer can still cause reported loss. Private memory grew about 117 MiB during capture, mostly in WebView; both AI workers exited after Stop. Physical unplug, ordinary-laptop battery/fan behavior, paid/authenticated cloud and fresh-machine WebView2 bootstrap were not tested. AI semantic accuracy remains a separate evaluation.

Full continuity report · CI guide

Build provenance

Source: 64aa24eda90919a226884eb59da34a32201396c2.

Windows checks · Cold Windows build

Runner: Windows 2025, image 20260925.250.1. Node v24.15.0, pnpm 11.25.0, rustc 1.98.1 (48a229cea 2026-09-01). No restored project/dependency caches or inference-model downloads. Installed native payload, icons and license/NOTICE verification passed on the hosted runner.

Installer: LectureRelay_0.3.3_x64-setup.exe, 16,177,076 bytes.

SHA-256:

8383AFFA98C757CA56E892833EC75F9D235A2B5D0451816718B6CE181E52E3EF

The .sha256 asset contains this checksum. build-metadata.json records toolchains, pinned inputs, source and installed payload hashes. Automatic Source code archives are for developers; use the EXE to install.

LectureRelay 0.3.2 - Reliability and readability preview

Install on Windows

Download LectureRelay_0.3.2_x64-setup.exe from Assets below, run it, then open LectureRelay. You do not need to build the project or install developer tools. Existing library data and preferences are preserved when upgrading.

This is an unsigned Windows x64 preview. It requires no LectureRelay account. Local model downloads happen separately in Settings; optional cloud services require the user's own provider and API key. This repository is currently private, so GitHub downloads require repository access.

Changes since 0.3.0

• Stronger contrast in Light and Dark themes, including secondary text, inputs, buttons and selected navigation.
• Safer recording and persistence: repeated/stale Stop requests cannot overwrite saved classes; concurrent file writes use independent temporary files; committed imports remain recoverable after database errors.
• Separate audio queue-drop and device-discontinuity diagnostics, with saved warnings visible in replay.
• More reliable note saving and duplicate-click handling; truncated cloud text is rejected instead of being marked complete.
• Resumable whole-class review with saved progress and bounded output-limit recovery.
• Delete permanently for courses in Trash requires typing DELETE. Free all storage… requires DELETE ALL and clearly lists its scope. Preferences, provider keys and the installed application remain.
• Bounded PDF reads and stable current-user installer verification.

Validation and limits

51 ordinary Rust tests, Clippy, TypeScript, production build and real React component checks passed. The installed 0.3.2 app completed a real 66.58-second local speech/translation run with seven bilingual segments, pause/resume, Stop, timestamp replay, theme switching and restart. All original data rows and 34 recording/model file hashes were preserved.

The prior 90-minute installed run belongs to 0.3.1 and is separate evidence. Device discontinuities, AI terminology/factual errors, ordinary-laptop power/noise, authenticated cloud workflows and fresh-Windows setup still need further validation. This is not an independent security audit or a guarantee of classroom accuracy.

Full engineering report · Prior 0.3.1 acceptance

Source commit: 1e183f241a09f6a4210cd0260c0dcd8c332db69d.

Installer size: 16,152,263 bytes.

SHA-256:

75750B72353BD99E421ED8E73B7EAE836B0F6C29F9DE7D26BB5877C3EAF77D70

The .sha256 asset contains the same checksum. The automatic Source code archives are for developers, not installation.

LectureRelay 0.3.0 - Local AI preview

Install

Download LectureRelay_0.3.0_x64-setup.exe from Assets and run it. No source build, Python, Ollama or developer tools are required. WebView2 is installed if missing. No LectureRelay account is required.

Local AI preview

• English speech: existing Nemotron Streaming EN 0.6B.
• Translation: Hy-MT2-1.8B Q4_K_M, with Qwen3.5-4B Q4_K_M as an alternative.
• Summary, whole-class review and Q&A: one shared Qwen3.5-4B download (attributed Unsloth conversion). Review accepts your instructions, processes every transcript section and saves a separate version, preserving manual notes.
• Separate translation/study provider choices; fresh text settings use Local. Existing explicit cloud choices are preserved. Local inference never falls back to a paid API.
• Global Quiet Mode gives local workers a shared budget of up to four logical CPUs. Remaining AI work after stopping has a separate status; workers are released after work or cancellation.
• Download only the models you need under Settings → Local AI. All three require about 4.3 GiB of disk space in addition to the app and recordings. The installer includes CPU runtimes, not weights. A GPU is not required.

Verified scope

32 normal Rust tests passed, plus Clippy, frontend type check/production build, formatting and repository checks. Opt-in tests exercised actual local models, download/hash/cancellation, worker authentication/crash cleanup, summary/review and source-linked Q&A. Ten native WebView checks and a 50.74-second system-audio → local speech → local translation → visible-caption/pause/resume/stop/replay flow passed using an isolated debug native app with production frontend.

The installer completed an isolated install with exit code 0. Installed version, executable payload, ASR/text runtime hashes, licenses and icons matched. The existing classroom database hash was unchanged. This does not establish full installed-release GUI acceptance.

Known limits

This is an unsigned preview. The small language test found terminology/negation errors; Qwen also made a summary interpretation error. These candidates are not certified as best or consistently accurate. Cold debug first translation took about 44 seconds; live latency is not yet certified. Ordinary-laptop noise/power and a new 90-minute full-app local-translation soak remain unverified. Prior WebView2 memory growth is not claimed fixed. No authenticated cloud calls were tested. Bergamot is a development comparison only. No CourseDude runtime is distributed; no image generation is included.

Setup · Evaluation and raw evidence · Provenance and licenses

Installer: 16,069,194 bytes. SHA-256:

E870496B2035F138BC77FDEA772BBF87FD69F09A7551A8F28CAE2DF6F96A7C58

LectureRelay 0.2.0 - Windows preview

LectureRelay 0.2.0 — Windows preview

Install and use

Download LectureRelay_0.2.0_x64-setup.exe below, run the setup wizard, and open LectureRelay. No source build, Node.js, Rust or Python is required. The Source code archives are for developers.

• Windows x64 (Intel/AMD); installs for the current Windows user.
• The installer downloads Microsoft WebView2 if it is missing.
• No LectureRelay account is required. This private repository's download access is separate from app login.
• For English captions, download the approximately 667 MiB speech model in Settings → Local AI, then select Local English in AI Providers. A GPU is not required.

Included

• Course background and glossary; microphone or system-audio recording.
• Local English speech recognition, live captions, pause/resume and Jump to Live.
• Transcript replay, timestamp navigation, search, corrections and exports.
• Media import, PDF slides, timestamped notes, bookmarks and chapters.
• Note versions, visible processing tasks, Light/Dark themes, global Quiet Mode and course Trash/restore.

Translation, generated notes and Q&A currently require an optional OpenAI/Groq provider with your own API key. Provider charges or limits apply. Cloud processing is disabled by default. Offline translation and local AI summaries are not included yet.

Verification and preview limits

The installer was built from the published v0.2.0 application source. Frontend production build, full regular Rust tests in the release profile (29 passed; 5 opt-in hardware/model/credential/fixture tests skipped), Clippy, formatting, repository hygiene and native runtime checks passed.

A fresh isolated installation exited successfully. Installed version, application payload, the speech worker, nine runtime DLLs, license notices and application/installer/uninstaller icons were verified. The existing classroom database was unchanged. These packaging checks do not establish full GUI or authenticated cloud acceptance for this exact installer.

This preview is unsigned. Authenticated cloud workflows and ordinary-laptop power/noise behavior remain unverified. A previous long classroom run found WebView2 memory growth; this release does not claim to resolve it.

Installer checksum

Size: 6,335,557 bytes (approximately 6.04 MiB).

SHA-256:

A2259600D646A2781D2931EE0517D492777C18EC0C95E362D038176D98C12493

The matching .sha256 file is included with the installer.