v0.7.8 — 听得见就得写得出：转写停摆当场报警 STT liveness watchdog & dictionary polish

转写停摆看得见 有声音进来、字幕却停了？状态栏当场亮「转写停滞」。判定很挑剔：五秒窗内累计一秒以上语音级声能才算「有声音」，攒满约一分钟、且 90 秒（按活动时间计）没有任何转写结果才报——Q&A 冷场、翻页静场、键盘咔哒、咖啡歇后恢复都不误报，暂停重来只认新证据。只提醒，绝不打断会议。双路收音下分声道判定，本地 whisper 引擎同样覆盖。
管线假死也报 音频统计整体断流 45 秒同样亮警，采集管线卡死不再无声无息。
单路无声提示 双路收音下某一路开场十分钟零声音、另一路正常时给中性提示（不是警报——本场不发言很正常），查一下输入设备即可。
词典边界收口 「mode (statistics)abc」这类粘连后缀不再误命中；非词字符结尾的短语不再吞屈折后缀（100%s 不再算 100%）；401(k)s 以词形变体回归；跨包词条碰撞从今有清单守卫测试，新碰撞必须显式过账。

文档站：https://mianaz.github.io/jargonslayer/

---

Transcription stalls are visible Audio flowing but no captions? The status bar now shows 转写停滞 on the spot. The predicate is deliberately picky: a 5-second window counts as "audio present" only with 1+ second of speech-level energy, and the warning needs about a minute of such windows plus 90 seconds of active time with zero results — Q&A lulls, silent slides, keyboard clicks, and post-break resumes don't false-fire; a resume starts from fresh evidence. Advisory only, never interrupts the meeting. Per-channel in dual capture; local whisper engines covered too.
Wedged pipelines too A 45-second gap in audio stats raises the same warning — a stuck capture pipeline is no longer silent.
Silent-channel hint In dual capture, a channel with zero audio ten minutes in (while the other is alive) gets a neutral hint (not an alarm — staying quiet is normal), suggesting a look at the input device.
Dictionary edges Glued suffixes like "mode (statistics)abc" no longer match; phrases ending in non-word characters no longer swallow inflections (100%s no longer counts as 100%); 401(k)s returns as a variant; cross-pack term collisions now have a ledger test — new collisions must be consciously booked.

Docs: https://mianaz.github.io/jargonslayer/

sha256

5186b922ebbe436c88754d37256cdb83f71f2948c0113e6365c2aeb81a287659  JargonSlayer-0.7.8.dmg
e696a6f3ec291832ed4838bea0306d999ad61e0a4e9300fd7fc8d7652e628414  jargonslayer-extension-0.7.8.zip

v0.7.7 — 撑得住：翻译断了看得见，长会不再越开越卡 Long-session hardening & update check

检查更新 设置里可手动查新版，后台每日静默核对一次；发现新版弹横幅，一键打开下载页（桌面版）。
翻译断线看得见 翻译通道在会议中途死掉也会亮「翻译失败」，不再只有全程零翻译才报；连续空响应同样计入；判定加了 60 秒地板，短暂抖动不再闪假警。
长会性能 90 分钟会议尾段逐条重算全表的三处热点改为增量缓存，未翻译计数不再随每条字幕全量扫描。
转录服务重连 本地转录服务闪断改为指数退避重连（1 秒起步，30 秒封顶），掉帧计入诊断；连挂五轮弹提示，此后每 5 分钟重报一次，始终不放弃重试。
定位刷新 README、标签页、首屏副标题、教程转向「捕获优先」：双路收音是主角，术语卡片是实时辅助。「全本地」只在默认本地引擎下声明，云端引擎为可选且界面有标注；双路收音的 macOS 26+ 门槛补齐。
修复 系统翻译准备超时会残留子进程的两处漏洞；切换历史会话后说话人计数串台；更新检查遇到 GitHub 限流不再记为已检查；静默检查失败不再显示「检查失败」。

文档站：https://mianaz.github.io/jargonslayer/

---

Update check Manual check in Settings plus a quiet daily check; a banner appears when a new version exists, one click opens the download page (desktop).
Translation death is visible A translation lane that dies mid-meeting now shows 翻译失败 — no longer only when zero translations landed all meeting; consecutive empty responses count too; a 60-second wall-clock floor keeps transient blips from flashing false alarms.
Long-session performance Three hot paths that recomputed the full transcript on every new line become incremental caches; the untranslated count no longer scans on every caption write.
Transcription reconnect A flapping local transcription service now backs off exponentially (1s to 30s cap), dropped frames are logged to diagnostics; after five failed cycles a notice appears and re-arms every 5 minutes — it never stops retrying.
Repositioning README, tab title, header subtitle, and tutorial move to capture-first: dual capture is the headline, jargon cards are live assist. "Fully local" is claimed only for the default local engines, cloud engines are labeled opt-in; the macOS 26+ floor for dual capture is stated.
Fixes Two holes where a timed-out translation prepare left a child process registered; speaker counts leaking across history-session switches; GitHub rate-limited update checks no longer count as completed; quiet-check failures no longer surface as 检查失败.

Docs: https://mianaz.github.io/jargonslayer/

sha256

5d1d2163d0a55069973267868b3d02f877e7433a406aeec4972115c9a4e2fc83  JargonSlayer-0.7.7.dmg
0778f8414d914e6f0bc19ad020709cee756c1fb461a41544c46400bc2f2122fd  jargonslayer-extension-0.7.7.zip

v0.7.6 — 听得全：自己和对方同时入记 Dual capture & UI polish

双路收音 麦克风与系统音频同时捕获，回声消除自动区分「我」和「对方」（macOS 26+，桌面版）。
三步引导 全新首次启动流程，演示一键直达。
状态栏诚实化 芯片显示实际运行模式（词典/模型/未配置）与隐私状态（本地/云端），手机端常驻。
手机界面收纳 工具按钮并入菜单，状态芯片折叠为 ⋯。
设置 文案口语化，主题色板可预览，Esc 关闭。
无障碍 卡片与状态播报适配屏幕阅读器，对话框焦点管理，触控目标加大。
学习中心 「开始复习」置顶，加载骨架屏。
构建 桌面 DMG 由 CI 构建并公证，扩展包随版本更新。
文档站：https://mianaz.github.io/jargonslayer/

---

Dual capture Mic and system audio captured simultaneously with echo cancellation; speakers auto-tagged as "me" vs "other" (macOS 26+, desktop).
Onboarding New 3-step first-run flow with one-tap demo.
Honest status bar Chip shows actual running mode (dictionary / model / not configured) and privacy state (local / cloud); always visible on mobile.
Mobile chrome Tool buttons moved into menu, status chips collapse to ⋯.
Settings Copy rewritten in plain language, theme palette preview, Esc to close.
Accessibility Card and status announcements for screen readers, dialog focus management, larger touch targets.
Study center "Start review" pinned to top, skeleton loading screen.
Build Desktop DMG built and notarized by CI; extension bundle updated with release.
Docs: https://mianaz.github.io/jargonslayer/

sha256

a9ae5797ae3d951c227b96921b5097473bf4da962c5d78715bcea6c9f5bad988  JargonSlayer-0.7.6.dmg
9c56b875ccf44546c3b7bfc4625ce9723dd437ded2a18654f1a74dfbd3727462  jargonslayer-extension-0.7.6.zip

v0.7.5 — 认得出：翻译不再无声死掉，麦克风不再录到静音 Recognition & recovery

中文

• 翻译可能在一整场会议里彻底消失：启动时握手失败一次就把翻译通道锁死，之后再无重试，只有最开头一条提示然后一片空白——同一天、同样设置的另一场却正常出双语。现在翻译通道自动重连，真连不回来时状态栏会写「翻译失败」，不再装作还在工作
• 桌面版选「麦克风」模式录到的全是静音。桌面版的系统识别走的是 CoreAudio 的系统输出捕获，根本没有麦克风通路；选了麦克风只是在输出捕获上贴了个麦克风的标签，音频帧在涨、峰值始终是零、什么都识别不出来。麦克风模式现在走本地模型。之前已经保存了错误配置的安装会在启动时自动修正
• 设置里的引擎卡片只写了引擎没写音频来源：选「系统识别」时来源还显示「麦克风」，底栏声称麦克风在录，实际走的是系统音频捕获
• 本地转录服务的端口（8765/8766）被上一次强退或第二个 app 实例占着时，启动直接崩出一段 Python 报错。现在提示是人话：哪个端口被占了、大概率是什么进程、怎么杀掉
• 设置和首次运行教程把音频来源当识别引擎来列，还在写「本地 Whisper」——实际本地模型也跑 Parakeet。术语已全部校正
• 系统识别引擎接受一组提示词来提高专业术语的转录准确率，以前平均分配给所有启用的词典包，一场生信报告大部分预算花在了金融和日常词汇上，真正的生信术语只分到大约十个名额。现在按当前会议的领域加权。个人生词本以前会独占整个预算，现在和词典包共享。这只提高已有词条被正确转录的概率，不等于新增了词条——湿实验/蛋白/免疫学包还没做
• 会议记录一键导出到 Obsidian 和 Notion
• 详见 https://mianaz.github.io/jargonslayer/

English

• Translation could die for an entire meeting: a single failed startup handshake latched the translation lane off with no retry, one toast at the very start and then nothing — while another session the same day with the same settings produced full bilingual output. The lane now retries on its own, and if recovery genuinely fails the status bar reads 翻译失败 instead of looking like it is still working
• Desktop microphone mode recorded silence. The desktop system recognizer is a CoreAudio system-output tap with no microphone path; selecting microphone started the output tap under a mic label — audio frames climbed, peak stayed at zero, nothing was recognized. Microphone mode now routes to the local model. Installs that were already saved in the broken state self-heal on launch
• The settings engine cards wrote the engine without its paired audio source, so picking 系统识别 while the source read 麦克风 left the bottom bar claiming the mic was live over a system tap
• A stale local transcription service on ports 8765/8766 — a force-quit leftover or a second app instance — crashed startup with a raw Python traceback. The error now names the port, the likely stale process, and the command to kill it
• Settings and the first-run tutorial listed audio sources as if they were recognizers, and still said 本地 Whisper for a local model that also runs Parakeet. Swept
• The system recognizer accepts a limited list of hint terms to improve specialist transcription. That list was split evenly across every enabled dictionary pack, so a bioinformatics talk spent most of its budget on finance and everyday vocabulary and got roughly ten slots of bioinformatics file formats. Hints are now weighted toward the domain the current meeting is actually about. A large personal glossary also used to consume the entire budget before any pack term was reached; glossary and packs now share it. This improves the odds a specialist term is transcribed at all — it does not add new dictionary entries; the wet-lab / protein / immunology pack is still to come
• One-click session export to Obsidian and Notion
• Details: https://mianaz.github.io/jargonslayer/

校验和 / Checksums

JargonSlayer-0.7.5.dmg  sha256: 613237934fa6a612b2156809d3dfb223e4d63c7da4346dd57707ccffda65bf14
jargonslayer-extension-0.7.5.zip  sha256: 101c9b930986e47dcd9fd7c32354474989b8f67a0e28cb89b3f50df6eabc59ad

v0.7.4 — 词典多义与查词重做 Word senses & lookup rework

中文

• 词典学会一词多义：讲癌症恶病质时 CAC 解释成 Cancer-Associated Cachexia，不再是获客成本；「IND … or NDA or BLA」里的 NDA 是新药申请；precision-recall 语境里的 PR 不再是 Pull Request。释义按当前会议的领域来选，领域不明时仍然给出可用的解释
• 划词改成先查词典：选中任意文字都直接查，命中的词典包和义项全部列出，不再只留一个赢家。真查不到时结合上下文解释并翻译，「加入生词本」退到次要位置
• 转录引擎按类型分组：本地 STT 模型、系统 STT 服务、第三方 STT 提供商。第三方服务没配 key 也常驻显示，标注「已配置 / 未配置」
• 音频来源独立成一个控件，会议进行中也能切换。系统/App 音频现在写明：录的是对方的声音，不含你自己的麦克风
• 网页版连不上自建转录服务时给一句人话提示，以前甩的是开发者才看得懂的 socket 报错
• 所有 API 密钥集中到「密钥」分类，此前 12 个填 key 的入口散在五个分区里。只改呈现方式，存储位置和 macOS 钥匙串托管都没动
• 详见 https://mianaz.github.io/jargonslayer/

English

• The dictionary now carries multiple senses per term: CAC resolves to Cancer-Associated Cachexia in a cachexia talk rather than 获客成本, NDA to New Drug Application inside "IND … or NDA or BLA", PR to precision-recall rather than Pull Request. Sense selection reads the meeting's domain, and an unknown domain still yields a usable card
• Selecting text now looks the word up first: every matching pack and sense is listed instead of a single winner, a genuine miss is explained in context and translated, and "add to glossary" drops to a secondary action
• The engine picker is grouped by kind — local STT models, system STT services, third-party providers. Unconfigured providers stay visible with a configured/not-configured hint instead of disappearing
• Audio source is its own control now, changeable mid-meeting. System/app audio states plainly that it captures the other party, not your microphone
• The web build shows a readable setup message when it cannot reach a self-hosted transcription service; it used to surface a raw socket error
• Every API key lives in one 密钥 settings category — 12 entry points were spread across five unrelated sections. Presentation only: storage and Keychain custody are unchanged
• Details: https://mianaz.github.io/jargonslayer/

校验和 / Checksums

JargonSlayer-0.7.4.dmg  sha256: 0c13bc706d270648107280842f9486e72e027e37fc74ae1bc41023036391a0ee
jargonslayer-extension-0.7.4.zip  sha256: 3b4d4781955fadf161b589497e59d43b5e0fea87812b7b1f599380537f8542f0

v0.7.3 — 字段测试修复 Field-fix train

字段测试修复版 · Field-fix release

First release driven by a real field-test cycle — six meeting transcripts, 43 dictionary cards, 12 of them wrong.

Fixed

• API keys are now per-provider. Switching the LLM provider previously transmitted the previous provider's key to the new provider's host. Keys are isolated per provider, migrated automatically, and stay in the macOS Keychain.
• Speaker rename. Renaming a speaker left the roster and the live picker stuck on the old name, so every later segment kept the stale label — in one session for 50 minutes.
• Jargon card context. Cards quoted whole 60-word ASR run-ons; they now narrow to the clause around the match.
• False positives. "in the red" no longer fires on "the red graph"; broken dictionary glosses (mean shipped as literally "1.") repaired at source with a build-time lint.
• Meeting-domain inference. "thank you for your attention" no longer produces an attention card in a biology talk, while precision/recall still fire in an ML talk.
• Wrong-sense cards. CAC now resolves to Cancer-Associated Cachexia rather than 获客成本, NDA to New Drug Application, PR to precision-recall.

Notes

• No binaries attached: DMG signing in CI is still pending the Apple-credentials decision, and this tag predates the release workflow. From v0.7.4 onward artifacts are built and attached automatically.

Commits

• Test-count badges 4745 -> 4761 (item-1 domain inference lane)
• feat(detect): infer meeting domains for common jargon
• v0.7.3 version stamp (8 sites + both lockfiles)
• fix: reject grammatical false positives for prior/mean/attention/recall
• Test-count badges 4718 -> 4745 (v0.7.3 field-fix train)
• fix: isolate LLM API keys by provider
• fix literal expression context collisions
• fix detection context and expression false positives
• fix: close dangling-colon gloss debt for four ml-stats terms
• fix: reject broken dictionary glosses and repair mean/precision/recall
• fix speaker rename leaving roster/latch stuck on the old name

v0.7.2 — 词典修复与诊断增强 Dictionary fixes & diagnostics

中文

• 词典修复：带括号的术语（如 stochastic gradient descent (SGD)、mode (statistics)）之前永远匹配不上，现在能正常识别
• 词典修复：只开专业词典包时不再丢词——同一个词在不同包里含义不同时（如 SAM），按你启用的包给解释；自己安装的词典包优先于内置
• 日志查看器覆盖三份本地日志（Whisper 服务、系统/App 音频、系统识别），可切换查看
• 新增导出诊断文件：把诊断信息存成文件（iOS 走系统分享面板），报问题更方便
• 浏览器插件新增复制诊断按钮，版本更新到 0.7.2，带上以上词典修复
• 隐私加固：本地日志确保不会记录会议原文
• 详见 https://mianaz.github.io/jargonslayer/

English

• Dictionary fix: terms with parenthetical qualifiers (e.g. "stochastic gradient descent (SGD)", "mode (statistics)") could never match before — now they do
• Dictionary fix: enabling only a specialist pack no longer drops entries — when a term means different things in different packs (e.g. SAM), you get your enabled pack's definition; user-installed packs take precedence over built-ins
• The Settings log viewer now covers all three local logs (Whisper service, system/app audio, on-device recognition), switchable
• New: export diagnostics to a file (via the system share sheet on iOS) — easier bug reports
• The browser extension gains a copy-diagnostics button and updates to 0.7.2, carrying the dictionary fixes
• Privacy hardening: local logs are guaranteed never to record meeting speech
• Details: https://mianaz.github.io/jargonslayer/

校验和 / Checksums

JargonSlayer_0.7.2_aarch64.dmg  sha256: 22aaaf3b2e986dd30ee71f923276a45e39f5d5e8a5dd20e7edeefcafef9f06ee
jargonslayer-extension-0.7.2.zip  sha256: 5eaf8257711452e0883d06c93f7458f824ac99a2728cfbd527f099f6040d48ee

v0.7.1 — 有道翻译与会后补全 Youdao translation & post-meeting gap-fill

中文

• 翻译引擎新增有道翻译：自带 appKey/appSecret，按量计费，适合中文场景
• 新增会后补全翻译：会议结束后一键把未翻译的段落补齐，走你配置的翻译引擎
• 导出跟随翻译：纪要导出带上已有译文，未翻译段落标出
• 存词改进：加入词典时记下该词在会议里实际出现的形式，复习时更好认
• 手机版界面：顶栏合并成单行，词条整词显示，控件尺寸统一
• 吉祥物 Bit 换了新造型，配色更暖，出场改成打招呼式
• 修复：卡片里的上下文长句不再被截断，触屏上也能看全
• 修复 iOS：导出改走系统分享面板，之前点导出没反应的问题解决了
• 稳定性：修复删除会话与保存中会议之间的竞态问题；详见 https://mianaz.github.io/jargonslayer/

English

• New translation engine: Youdao — bring your own appKey/appSecret, pay per use, strong for Chinese
• Post-meeting translation gap-fill: after a meeting ends, fill in untranslated segments with one click via your configured engine
• Exports follow translations: summary exports carry existing translations; untranslated segments are marked
• Word saving: entries added to your dictionary record the exact form that appeared in the meeting
• Mobile UI: single-row header, full-word chips, unified control sizing
• Bit got a new look — new body, warmer palette, greeting-style entrance
• Fix: long context sentences in cards are no longer cut off (fully visible on touch too)
• iOS fix: exports now open the system share sheet (tapping export previously did nothing)
• Stability: fixed races between session deletion and in-flight meeting saves; details: https://mianaz.github.io/jargonslayer/

浏览器插件本次无改动，附件与 v0.6.0 相同 / Extension unchanged this release (zip identical to v0.6.0).

校验和 / Checksums

JargonSlayer_0.7.1_aarch64.dmg  sha256: a76841fa5444b4d17026d870dcc1bc5d59047b83150073fd342287b05ec1dbff
jargonslayer-extension-0.6.0.zip  sha256: 3f51a97d5f4032bbccdf4eabd6d5cc2e69e017fbf12186d296ae72c2cdce06a1

v0.6.0 — 词典包与系统翻译 Dictionary packs & on-device translation

中文

• 设置里可从官方目录一键安装专业词典包（医学、生物医药、AI/ML、法律、金融商务、工程计算机，6 个包共 4300+ 词条）；也能从文件或 GitHub 链接导入第三方包；已装的包每天自动检查更新
• 内置词典新增三个通用包共 271 条：新兴用法（harness、agent、RAG…）、钱与福利（保险/薪酬/股权）、日常口语与俗语（寒暄、委婉说法、体育比喻）
• 多义词按语境排序——同一个词在不同领域含义不同（如 EMT），自动把当前会议最可能的解释排前面
• 翻译引擎新增 Apple 系统翻译（macOS 26+），本机运行，不需要 Key，文本不出设备
• 语音识别新增 ElevenLabs 云端引擎，自带 Key 按量计费
• API Key 存进 macOS 钥匙串（首次启动自动迁移）；设置新增网络代理字段（也跟随系统代理）；用量统计记在本机随时可查，不上传服务器
• 开会一两分钟后自动推断会议主题背景，检测更贴合语境；背景可点击编辑
• 模型下载支持转后台或取消；浏览器插件同样支持安装词典包
• 详见 https://mianaz.github.io/jargonslayer/

English

• Install specialized dictionary packs from the official catalog in Settings (medicine, biopharma, AI/ML, law, finance, engineering — 6 packs, 4,300+ terms); import third-party packs from a file or GitHub link; installed packs auto-check for updates daily
• Three new built-in general packs (271 terms): emerging usage (harness, agent, RAG…), money & benefits (insurance, compensation, equity), everyday idioms & slang
• Multi-sense terms sort by context — when a word means different things in different fields (e.g. EMT), the most likely definition for the current meeting comes first
• New translation engine: Apple on-device translation (macOS 26+), runs locally, no API key, text stays on your machine
• New speech recognition engine: ElevenLabs cloud STT, bring your own key, pay per use
• API keys stored in macOS Keychain (auto-migrated on first launch); network proxy field added in Settings (also follows system proxy); local usage stats for speech and AI calls viewable in Settings, nothing uploaded
• Meeting context auto-detected after ~1–2 minutes for better detection accuracy; click to edit
• Model downloads can run in the background or be canceled; browser extension also supports dictionary pack installs
• Details: https://mianaz.github.io/jargonslayer/

校验和 / Checksums

JargonSlayer_0.6.0_aarch64.dmg  sha256: a275baf0ebba029caa134e1430dbe63036de84c3bc398efba6e5cf5e32eb014d
jargonslayer-extension-0.6.0.zip  sha256: 3f51a97d5f4032bbccdf4eabd6d5cc2e69e017fbf12186d296ae72c2cdce06a1

v0.5.0 — 模式优先 · Mode-First

中文

• 开会前先选「怎么听」：麦克风、浏览器标签页、系统与 App 音频、导入文件，引擎自动匹配
• 说话人标注：逐句或批量指定，可应用到本句及之后，实时显示当前说话人
• AI 转写修订（审阅制）：逐段对比，接受才生效，之后自动重新翻译
• 崩溃恢复：开会时自动存草稿，刷新或崩溃后一键回到历史记录
• 划词解释后台生成：弹窗关了也不丢，照样跑完并存进卡片
• 翻译引擎可选 LLM 或 Chrome 内置本地翻译（Chrome 138+，免 Key）；词汇表给识别加偏置，术语更少写错
• 导出与联动：docx 导出、AnkiConnect 同步、自定义命名词包、卡片就地编辑
• 在线体验新增 Soniox 云端识别限量试用（免 Key，每人每天 3 段、单次约 10 分钟、先到先得）；标签页音频在线回归
• 详见 https://mianaz.github.io/jargonslayer/

English

• Pick how you listen (mic / browser tab / system audio / import) — the engine picks itself
• Speaker tools: per-segment & bulk assign, apply-to-rest, live current-speaker
• Review-gated AI transcript correction with auto re-translation on accept
• Crash recovery: live meeting drafts, one-click restore after refresh/crash
• Background selection lookups — close the popover, cards still land
• Translation providers (LLM or Chrome's built-in, 138+); glossary-biased recognition
• docx export, AnkiConnect sync, named custom packs, inline card editing
• Hosted preview: free limited Soniox cloud STT trial (no key; 3/day per person, ~10 min each) + tab audio returns
• Details: https://mianaz.github.io/jargonslayer/

校验和 / Checksums

JargonSlayer_0.5.0_aarch64.dmg  sha256: fee39164305359c13490bb5149539c84cd281f0c57801c72df03979a91d1c4a9
jargonslayer-extension-0.5.0.zip  sha256: 564ca38d52d2d2f6eddfb3f7d80239dd7d7a691ee384e17c982897844249ad82