Release History
Every user-visible change, newest first. Versions cover the web app, the macOS desktop app, the iOS beta, and the Chrome extension together unless a change names a specific platform. Release downloads live on GitHub releases; what's planned next is on the Features page.
v0.7
v0.7.8
2026-08-05Latest- Transcription stall warning: audio flowing but no captions now raises 转写停滞 on the spot — the verdict needs about a minute of speech-level energy with zero results plus a 90-second floor in active time, so Q&A lulls, keystrokes and post-break resumes never false-fire. Advisory only, never interrupts; per-channel in dual capture, local whisper engines covered.
- 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 (other channel alive) gets a neutral hint — staying quiet is normal, a muted mic is worth a look.
- Dictionary edges: glued suffixes ("mode (statistics)abc") no longer match; 100%s no longer counts as 100%; 401(k)s returns as a variant; cross-pack term collisions now have a ledger test.
v0.7.7
2026-08-04- Update check: manual check in Settings plus a quiet daily check; a banner opens the download page when a new version exists (desktop).
- Translation death visible: a lane dying mid-meeting now shows 翻译失败, not only after a fully silent meeting; a 60s floor prevents false alarms.
- Long-session performance: hot paths that recomputed the full transcript per line become incremental caches.
- Transcription reconnect: exponential backoff (1s→30s), dropped frames logged, a notice after five failed cycles, retries never stop.
- Repositioning: capture-first copy — dual capture is the headline, jargon cards are live assist; "fully local" claimed only for the default local engines.
- Fixes: translate-prepare timeouts leaving a child process; speaker counts leaking across history loads; rate-limited update checks miscounted as completed.
v0.7.6
2026-08-04- Dual capture: mic + system audio with echo cancellation, auto me/other tagging (macOS 26+ desktop).
- New 3-step onboarding with one-tap demo.
- Status bar shows the actual running mode and privacy posture, always visible on mobile.
- Mobile chrome folded.
- Plain-language settings + theme swatches.
- Accessibility: screen-reader announcements, dialog focus, larger touch targets.
- Study center leads with Start review.
v0.7.5
2026-08-02- Translation no longer dies silently: a single failed startup handshake used to latch the translation lane off for an entire meeting, with one notice at the very start and then nothing — while another session the same day on 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 system-output tap with no microphone path, so selecting 麦克风 started the output tap under a mic label and recognized nothing. Microphone mode now routes to the local model, and installs already saved in the broken state repair themselves on launch.
- Capture source can no longer be mislabeled: 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.
- Recognition hints follow the meeting's domain: the system recognizer accepts a limited list of hint terms to improve specialist transcription, and that list was split evenly across every enabled dictionary pack — a bioinformatics talk spent most of its budget on finance and everyday vocabulary. Hints are now weighted toward the domain the meeting is actually about, and a large personal glossary no longer consumes the whole budget before any pack term is reached. This improves the odds a specialist term is transcribed at all; it does not add new dictionary entries.
- One-click session export to Obsidian and Notion.
- Readable port conflicts: a leftover local transcription service on ports 8765/8766 — a force-quit remnant or a second copy of the app — used to crash startup with a raw Python traceback. The error now names the port, the likely process, and the command to end it.
- Terminology sweep: Settings and the first-run tutorial no longer list audio sources as if they were recognizers, and no longer say 本地 Whisper for a local model that also runs Parakeet.
v0.7.4
August 2026- Domain-aware multi-sense definitions: a headword with more than one definition now ranks them by the meeting's domain — CAC resolves to Cancer-Associated Cachexia in a cachexia talk rather than Customer Acquisition Cost, NDA to New Drug Application inside "IND … or NDA or BLA," and PR to precision-recall rather than Pull Request in an ML talk; an unclear domain still returns a usable definition.
- 划词 lookup rework: selecting text now runs a dictionary lookup first — every matching pack and sense is listed instead of a single winner — and a genuine miss is explained in context and translated;
加入我的词典drops to a secondary action. - Engine picker grouped by kind: 本地 STT 模型 / 系统 STT 服务 / 第三方 STT 提供商 replace one flat list; unconfigured third-party providers stay visible with a 已配置/未配置 hint instead of disappearing.
- Audio source as its own control: changeable mid-meeting now instead of only before one starts; system/app audio states plainly that it captures the other party, never your microphone.
- One 密钥 settings category: every API key — LLM provider, transcription, translation, diarization token — now lives under a single 设置 → 密钥 screen instead of five scattered entry points; storage and Keychain custody are unchanged.
- Readable connection errors: the web build shows a plain-language setup message when it can't reach a self-hosted transcription service, instead of a raw socket error.
v0.7.3
2026-07-30Fix train- Per-provider API keys: switching LLM providers no longer risks sending the previous provider's key to the new provider's host — each provider's key is now stored and sent in isolation, migrated automatically, still Keychain-backed on desktop.
- Speaker rename fix: renaming a speaker now updates the roster and the live speaker picker immediately; it used to stick on the old name for the rest of the meeting.
- Tighter card context: an expression card's quoted context is now the clause around the match, not the full 60-word ASR run-on it came from.
- Fewer false positives, cleaner glosses: in the red no longer fires on unrelated phrases like the red graph, and broken built-in glosses (
meanshipping as literally "1.") are fixed at the source, with a new build-time lint to catch the next one. - Sharper meeting-domain inference: "thank you for your attention" no longer produces an attention card in a biology talk, while precision/recall still fire correctly in an ML talk.
v0.7.2
2026-07-28Fix train- Dictionary matching fixes: a headword ending in a parenthetical qualifier — stochastic gradient descent (SGD), mode (statistics) — now matches in the transcript (it never could before); enabling only a specialist pack (e.g. 生信/统计) no longer silently drops an entry whose name collides with a business-pack term, and each enabled pack's own definition surfaces; an installed dictionary pack now wins over a built-in one when both define the same term.
- Diagnostics for beta feedback: Settings' 诊断信息 log viewer now covers all three local logs (Whisper 服务 / 系统/App 音频 / 系统识别) via a source picker; a new 导出诊断文件 button saves the diagnostics bundle as a file (opens the system share sheet on iOS); the Chrome extension side panel gets its first 复制诊断 button.
- Local logs never carry meeting speech: a hardening fix in the desktop speech-recognition log writer closes the one remaining path where a malformed line could leak raw speech into a local log file.
- Chrome extension updated to 0.7.2: carries the dictionary reliability fixes above; the previous release was pinned at 0.6.0.
v0.7.1
2026-07-28- 有道翻译 Youdao translation engine (BYOK): another translation-engine option alongside the on-device translators, the LLM, and DeepL — bring your own 应用ID/应用密钥, billed per use; native-app only (desktop/iOS) since Youdao's API has no cross-origin support for the web app.
- 会后补全翻译 post-meeting gap-fill: a one-click toolbar button after the meeting stops re-translates any segment that was missed, through whichever translation engine is configured.
- Exports follow translations: Markdown, docx, and copy-to-clipboard exports now carry existing translations instead of only the ones an LLM summary happened to produce; if segments are still untranslated, you're offered a gap-fill-then-export option, with a 未翻译 marker for whatever remains.
- 存词 glossary surface capture: saving a card to your glossary now keeps the exact wording it appeared as in that meeting, not just the dictionary's canonical form.
- Mobile UI polish: a single-row header on phone widths, full-word status chips instead of abbreviations, and unified control sizing.
- Bit refresh: a redrawn body, a warmer color palette, and a greeting-style entrance animation for the status-line mascot.
- Fix: the source-sentence quote in an expanded expression card no longer clips at two lines.
- iOS fix: exports (Markdown, docx, backups, Cornell notes) now hand the file to the system share sheet — previously tapping an export button on iOS did nothing.
v0.6
v0.6.0
2026-07-24- Dictionary packs — 17 packs, 1,000+ entries: the built-in dictionary grew to 17 packs with multi-sense ranking that sorts definitions by meeting context. Variant matching (85 entries) catches inflected and alternate forms — say keep the ball rolling and the dictionary's get the ball rolling still hits.
- Apple on-device translation on iOS: the Apple Translation engine, previously desktop-only, now runs on iOS 26+ as well — fully on-device, no API key, text never leaves the device.
- iOS enters TestFlight beta: JargonSlayer is now available on iOS through TestFlight (by invite, iOS 26+). On-device system speech recognition and Apple Translation both run locally on the phone.
- Mobile main-screen redesign: the transcript now fills the phone screen, with explanation cards in a swipeable bottom strip — newest on the left, tap a highlighted term to jump to its card.
- Dictionary mode is now the default: new installs start in dictionary mode (the always-on, fully offline detection layer); AI mode moves to opt-in beta.
- 0.6.1 / 0.6.2 follow-ups: iOS build surfaced in the titlebar, and Apple on-device translation reaching every desktop and iOS surface it applies to.
v0.5
v0.5.1
July 2026- Appearance expansion: two more built-in themes — 像素 (8-bit quest log) and 笔记 (classroom notebook) — bringing the total to nine; a custom theme editor with live contrast hints and JSON import/export; separate UI and monospace font pickers; an optional frosted-glass overlay; and independent global/transcript type-size controls.
- Bit's wardrobe: the status-line mascot dresses to match each built-in theme, with manual pinning and sleep/celebration animations.
v0.5.0
2026-07-20- Mode-first UI: before a meeting, pick how you're listening (microphone / tab audio / system-app audio / an imported file) instead of an engine directly; the app matches the right engine to your platform and configuration automatically, and onboarding was rewritten to match.
- Speaker tools: per-line or batch speaker assignment, "apply to this and following," and a live "current speaker" tag; manually assigned lines are locked and survive later automatic diarization updates.
- AI transcript correction (reviewed): one click after the meeting has the AI read the transcript and flag recognition errors, listing the original and a suggested fix segment by segment — nothing changes until you accept it, and accepted segments re-translate automatically.
- Crash recovery: the meeting is auto-drafted in real time; after a page refresh or a browser crash, history offers a one-click 「恢复到历史记录」 restore.
- Background term lookup: a selection lookup's explanation keeps generating in the background and lands in the card list even if you close the popup.
- Exports and integrations: a new docx export; AnkiConnect syncs new cards into your local Anki in one click with dedup; the personal glossary can split into multiple named custom packs.
- Translation engine choice: bilingual subtitles and full-transcript translation can use the browser's built-in on-device translator (Chrome 138+, no key required) instead of the LLM; the desktop and iOS apps add Apple's on-device translation (macOS 26+ / iOS 26+), another local, key-free option.
- Deepgram cloud recognition (BYOK, English): a new cloud STT engine option for English-only meetings.
- Hosted preview's limited trial: a built-in Soniox cloud-recognition trial (no key, up to 3 clips per person per day, ~10 minutes max per clip, shared first-come quota); tab audio is back in the hosted preview on the same trial quota.
v0.4 series
v0.4.0 – v0.4.6
2026-07-12 → 2026-07-19- Desktop app (Tauri, macOS, v0.4.0): self-managed local Whisper — consent-gated wizard, model picker + live model switching, one-click diarization add-on. Signed, notarized DMG on releases.
- On-device browser recognition: Chrome 139+ prefers local speech recognition, automatic cloud fallback.
- Soniox realtime engine (experimental, BYOK): aimed at code-switched zh-en; stays experimental until it clears the in-house benchmark.
- System/App audio (desktop, macOS 14.4+): CoreAudio process-tap capture of what the Mac is playing, replacing desktop's old tab-audio slot.
- 系统识别 System Speech (desktop, macOS 26+, v0.4.3): zero-install engine on macOS's native SpeechAnalyzer — no model download, no Python; audio never leaves the capturing process.
- 英文加速 Parakeet (desktop, Apple Silicon, v0.4.4): opt-in, MLX-accelerated English engine in the model picker — far faster than Whisper (~100x realtime in batch), no realtime speaker diarization yet.
- Chrome extension (Lite): MV3 side panel with live mic capture, the built-in dictionary, and on-device translation.
- Desktop shell features: background task center, engine dropdown in the status bar, first-run onboarding (OpenRouter OAuth + diarization token), and a release update check.
- AI transparency panel + domain dictionary packs (v0.4.6): a per-task AI status panel (detection / explanation / translation / reports) with model, routing, and call/failure/QC-drop counts plus a status-line chip; three new built-in dictionary packs — statistics, machine learning, bioinformatics (~226 terms); and a detection-quality fix so ordinary sentences are no longer mislabeled as jargon.
v0.3
v0.3.0
2026-07-10- The learning release: glossary-only mastery grew into a first-class learn-set — known-word suppression on live cards plus a due-driven review loop (SM-2-lite scheduling) under
/review → 到期复习. - Unified import hub + background task center: transcript, recording, and URL imports in one place, with job progress tracking.
- Simple/advanced settings: settings split into a simple tier for daily use and an advanced tier for the full surface.
- Four new themes joined the original terminal pair.
- STT reliability: a VAD supervisor fix closed a transcript-loss path in live recognition.
- Diagnostics: the first redacted diagnostics bundle for bug reports.
Early Versions
v0.2.0 – v0.2.3
2026-07-06 → 2026-07-07- Content ingestion phase 2 (v0.2.0): audio, video, and URL import; bilingual transcription; OpenRouter one-click OAuth.
- v0.2.1 – v0.2.3: copy polish, server-route hardening after an adversarial review pass, and a translate-route model-compatibility fix.
v0.1.0
2026-07-06- First cut: live transcription with browser recognition, the built-in dictionary's instant detection floor, LLM explanation cards, bilingual summary, and local session history.