Cicada — the open-source menubar app at github.com/wquguru/cicada — is climbing GitHub trending lists because it solves an annoying problem: live bilingual subtitles and meeting notes for any audio on your computer, without inviting a bot into the call. The project's README (Apache-2.0) describes capture from Zoom, Teams, Feishu, Google Meet, Discord, X Spaces, YouTube, Twitch, and browser tabs, then structured notes with topics, conclusions, to-dos, and glossaries.
For AI builders, Cicada is a reference stack for voice-agent input: on-device ASR on Apple silicon, pluggable translation backends, and a privacy posture that keeps recordings local unless you turn on sharing.
TL;DR
| Question | Answer |
|---|---|
| License | Apache-2.0 |
| Platforms | macOS 14.4+, Windows 10 2004+, Ubuntu 22.04/24.04, Android 10+ |
| Bot-free? | Yes — taps system audio, not a meeting participant |
| On-device ASR | Apple silicon Macs (Intel/cloud needs a key) |
| Interpreting | Hold right Option (Mac) for translated voice into the call |
| Records | Local LiveSubs folder; export Markdown, Word, SRT |
Why developers noticed it now
The repository was created November 1, 2026 and quickly accumulated about 2,100 GitHub stars in the first week — enough to land on trending feeds with a mature feature list (660+ commits on the default branch when we fetched the page). That is unusual for a solo-maintained Swift/SwiftUI Mac core with parallel Windows, Linux, and Android trees.
The pitch is not "another Whisper wrapper." Cicada bundles:
- Live notes with fishbone diagrams and glossary extraction while the meeting runs.
- Language detection within seconds, with subtitles that fade when nobody speaks.
- Hold-to-talk interpreting that routes translated speech back through the meeting mic via BlackHole 2ch on Mac (one-click driver install).
- Speaker diarization using a bundled WeSpeaker voiceprint model on device.
- Share links with expiry, optional passwords, and a sensitive-content scan before upload.
That combination maps cleanly to teams prototyping voice agents or human-in-the-loop translation without standing up a full WebRTC pipeline.
Architecture: what runs where
Cicada open-source live subtitles bridge system audio into bilingual meeting notes
The README's Models table is the cheat sheet:
| Feature | With Ling API key | Without a key (Apple silicon) |
|---|---|---|
| Recognition | Apple on-device speech stack | Same on Apple silicon |
| Sentence translation | Ling-3.1-flash (configurable) | Google Translate → Apple Translation on macOS 26 |
| Partial-line translation | Google Translate streaming | Shows source text until sentence completes |
| Live AI notes | Ling flash models | Not available without a key |
| Interpreting TTS | Ming TTS or cloned voice sample | macOS built-in voices only |
Privacy split: Apple silicon keeps audio capture and recognition on laptop for the default path. Only text (and short hold-to-talk clips if you configure neural voices) leaves for translation or notes. Windows, Linux, and Android push audio segments to whichever cloud ASR you configure — OpenRouter, ZenMux, self-hosted vLLM, Whisper, or Qwen3-ASR endpoints are explicitly named in the README as compatible examples.
Cicada documents no analytics, no tracking, and a short allowlist of network hosts (translation, your AI vendor, BlackHole download with SHA-256 check, GitHub update ping).
Permissions and the "no bot" tradeoff
Mac users grant capabilities lazily:
| Permission | Purpose if granted | If denied |
|---|---|---|
| Screen & System Audio Recording | Hear meeting audio | No subtitles or notes |
| Microphone | Room audio and interpreting | No in-room or talk-key modes |
| Input Monitoring | Detect right Option hold | Pick another shortcut in Settings |
| Admin password (once) | Install BlackHole 2ch virtual mic | Remote side cannot hear interpreting |
Linux users should note the README warning: notes windows remain visible in screen shares, unlike Mac's default hidden capture. Android cannot record internal call audio — the workaround is speakerphone + In person mode.
These constraints matter when you compare Cicada to cloud notetakers like Medium's AI notebook MCP or enterprise Copilot meeting features in our Microsoft Copilot use-case guide.
Practical setup paths
Path A — Apple silicon, zero key. Install from Releases, grant system audio, use on-device recognition plus Google or Apple translation fallbacks. You get subtitles and local records, not full AI note polish.
Path B — Ling key. Register on Ant Group's Ling platform, paste sk- key into Settings, unlock live notes, neural interpreting, and voice cloning from an 8–30 second sample.
Path C — Bring your own model. Settings supports OpenRouter, ZenMux, OpenAI-compatible gateways, including LAN Ollama or SGLang. README warns Windows/Linux/Android recognition latency runs 1–2 seconds behind Ling when you switch.
For streaming ASR research complementary to this product layer, see Netease Youdao's new Confucius4-R2T2 stable-prefix model — Cicada does not ship R2T2, but voice-agent builders evaluating non-revising transcripts should read both posts.
Cicada on Apple silicon can keep speech recognition on-device for live meeting notes
Features power users care about
Word list and live correction. Paste canonical spellings (people, codenames, product names). Cicada fixes sound-alike errors inline and supports batch Edit line fixes across a session with undo.
Re-listen pass on macOS 26. Apple silicon builds can re-run on-device recognition 15–30 seconds after each line to fix boundary typos — a nice match for semantic caching thinkers optimizing double-pass workflows.
Imports and exports. Drag local audio/video through File → Import Recording or Video to build bilingual records after the fact. Export paths include Markdown, Word, and SRT for caption pipelines.
Roadmap honesty. The README lists missing continuous capture of your own unmuted speech in meetings as a top priority — today, "you" in the transcript mostly means hold-to-talk segments.
Comparison anchors
| Tool | Angle |
|---|---|
| Microsoft VibeVoice | Research models for ASR/TTS; not a menubar meeting product |
| Google AI Edge Foresight | Google demo of local notes; not cross-platform open source |
| Hugging Face speech-to-speech guide | How to wire STT→LLM→TTS for agents |
| Shizuku duplex voice | Research-grade full-duplex; not a shipping notes app |
Cicada sits closest to daily driver productivity: fewer benchmark charts, more permissions dialogs and hold-key muscle memory.
Security and procurement checklist
Teams evaluating Cicada for regulated calls should read the README's sharing scanner (flags keys, unreleased financials) and remember Share… uploads note content to Cicada's link host — treat it like any SaaS export.
For agent builders, the relevant lesson is data minimization: Cicada proves many meeting-AI features only need recognized text leaving the device, not raw multi-hour audio — similar to how sleep EEG consent posts discuss on-device boundaries.
If you deploy autonomous agents that summarize meetings programmatically, keep human review on outbound actions — AgentBeam is one guardrail layer when agents can send mail or patch tickets based on transcript triggers.
Android and hybrid meeting caveats
Android builds cannot tap in-call audio for most meeting apps — OS policy, not a Cicada bug. The documented pattern is speakerphone + In person mode, which bleeds audio to the room and reduces translation quality. For confidential calls, Android may be the wrong client even though the APK exists.
Hybrid rooms use the same In person microphone path on Mac: online audio is tagged Online, room audio In room, merged into one record. macOS 14–15 behaved slightly differently (mic-only capture in some builds); read the version notes in the README before you demo to executives.
Troubleshooting installers teams hit
Mac "Open Anyway" gate. Cicada admits it is not Apple-notarized yet. First launch requires Privacy & Security → Open Anyway. Security teams should plan for that friction until the roadmap notarization item ships.
BlackHole install prompt. Hold-to-talk interpreting needs the virtual mic driver. Without admin approval once, remote participants will hear your raw language, not translated speech.
Linux screen-share visibility. Notes windows show up in Zoom shares unless you toggle the setting — unlike Mac's default hide-from-capture behavior.
Windows SmartScreen. README tells users to click More info → Run anyway — factor that into company laptop policies.
Cost modeling with third-party keys
If you skip Ling and route through OpenRouter or a self-hosted vLLM ASR endpoint, you trade ¥-priced minutes for token-metered cloud bills. The README is explicit that non-Ling backends add 1–2 seconds of subtitle lag on Windows/Linux/Android — acceptable for internal notes, painful for live interpreting.
Pair Cicada with our KV cache explainer when you run local inference: streaming meetings generate long contexts if you feed full transcripts back into summarization loops.
How to try it today
- Download the latest Release asset for your OS from github.com/wquguru/cicada/releases.
- Verify SHA-256 against
checksums.txtin the release (commands are in the README). - Walk through onboarding sample audio before granting system audio on a live customer call.
- Pick AI Service in Settings; start without a key on Apple silicon if you only need subtitles.
- Read
docs/ROADMAP.mdfor features still in flight before you promise executives "full auto notes."
Bottom line
Cicada is real software, not a weekend demo: multi-platform clients, on-device models where Apple allows, honest key requirements elsewhere, and a no-bot capture model compliance teams may prefer. It trended because practitioners want live AI notes without another meeting bot name on the participant list.
We will update if the project ships code signing/notarization (on the public roadmap) or adds continuous self-speech capture in meetings.
Related reading
- Confucius4-R2T2 stable-prefix streaming ASR
- Microsoft VibeVoice open voice models
- Hugging Face speech-to-speech voice agents
- Semantic caching for LLM apps
- Shizuku full-duplex voice research
- Medium AI notebook MCP for private libraries
Version and permission details follow the Cicada README as of November 10, 2026.