Every call. Every stream. No language barrier.
Understand, speak back, and remember everything. Live meeting notes, bilingual transcripts, subtitles and interpreting on Mac, Windows, Linux and Android.
In meetings or livestreams, Cicada gives you live bilingual subtitles and writes structured meeting notes as it listens. When it's your turn to speak, hold the talk key (right ⌥ on Mac, right Ctrl on Windows, customizable) and talk—the other side hears your translated voice right in the call. Cicada listens to system audio directly without sending bots into the meeting, so no one else notices a thing. On Apple silicon Macs, all speech recognition runs on device, so audio never leaves your computer.
Where it works: Anything that plays sound on your computer—Tencent Meeting, Feishu, DingTalk, WeCom, Zoom, Teams, Google Meet, Webex, Discord, and Telegram voice chats, plus X Spaces, YouTube, Twitch, Bilibili Live, podcasts, and web pages. On phones, the system blocks recording audio from call and meeting apps—put it on speaker and use In person instead.
| Platform | Requirements | Notes |
|---|---|---|
| Mac | macOS 14.4+, Apple silicon / Intel | Full feature set. On-device recognition on Apple silicon (works without a key); Intel needs a Ling key (speech recognized by Ming) |
| Windows | Windows 10 2004+, x64 / ARM64 | Needs an AI service key (Ling, OpenRouter, ZenMux, etc.); see windows/ |
| Linux | Ubuntu 22.04 / 24.04 desktop | Needs an AI service key; the notes window is visible in screen sharing; see linux/ |
| Android | Android 10+ | Needs an AI service key; subtitles for videos and streams, face-to-face interpreting; see android/ |
Installers are available on the download page and Releases. All four platforms share the same version number and meeting record format. The sections below use the Mac app to walk through the features.
As you listen, Cicada organizes a summary, topics and key points, conclusions, to-dos, a fishbone diagram, and a glossary of names and terms. Click any conclusion or to-do to jump straight back to the original lines. Press ⌘F to search—even misrecognized homophones sho
8000
w up.
With an AI service connected, misheard names and terms are fixed on the spot: homophones are corrected right away, while unfamiliar names are asked about once and fixed automatically once you confirm. You can paste common terms into Word list in Settings ahead of time. After the meeting, select Edit line… to batch-fix the same misheard word across the whole session, with full undo support.
Cicada detects the spoken language within seconds and translates it into your system language; press ⌃⌥S to switch display modes. The subtitle bar only appears when someone speaks and turns see-through under your pointer so it stays out of the way. On macOS 26 with Apple silicon, Cicada also re-listens on device to refine each line 15 to 30 seconds later, fixing typos and missed words.
Click Turn on mic to capture people speaking in the room. Hybrid meetings use this same button: computer audio is marked Online and microphone audio is marked In room, combined in a single record (on macOS 14 and 15, turning on the mic listens only to the room). On speakers, Cicada automatically deduplicates audio; with headphones on, both sides can talk at once and both are kept. Windows and Linux follow the same rules.
On a Mac, the small arrow beside the button picks which microphone to use (an iPhone or a conference mic, say) without changing the system default. What the mic hears only goes into the notes, never into the meeting; for the other side to hear you, use simultaneous interpreting below.
Hold the right ⌥ key and speak so the other side hears your translated voice in the meeting app.
On a Mac, you can also turn on Hands-free in Settings › Interpreting: just start talking while listening to a meeting, the meeting audio automatically ducks, and your translated voice goes out after a 1-second pause. It pauses while the mic is on, and we recommend switching back to the talk key when others are nearby so their voices aren't mistaken for yours.
Turning on interpreting switches the system microphone to BlackHole 2ch, a virtual driver you can download and install with one click. With a Ling API key, you can use natural voices or clone your own voice from an 8 to 30 second recording; without a key, you can still use built-in macOS voices.
Bilingual transcripts, notes, and raw audio recordings are saved on your device. Click any line to play from that point while the text highlights and scrolls along. The on-device voiceprint model automatically separates speakers (and fills in their names when a Ling key is set), and you can rename or merge them anytime if someone is misidentified. Right now "you" in the record only covers what you say while holding the talk key; capturing what you say directly in the meeting is on the roadmap.
After a meeting, you can generate notes retroactively, export to Markdown, Word, and SRT subtitles, or transcribe local audio and video files via File › Import Recording or Video…. Turn on Settings › General › Save to folder and each meeting is also automatically saved as Markdown (with frontmatter at the top for agents) into an Obsidian vault or another folder, updating automatically whenever you rename it or regenerate its notes.
Click Share… to generate a link to the notes (optionally including the full bilingual transcript) with an expiration time and password; before uploading, it automatically blocks keys, unpublished company info, and other sensitive content.
Apple silicon Macs: Recording and speech recognition run on your computer; audio never leaves it. Only the recognized text is sent out (for translation and meeting notes). Exceptions: the lines you speak during interpreting with a Ling key configured, clicking Use Ling for Now, or switching to a self-hosted service in Settings › AI Service will send audio to the corresponding service. Without an API key, live subtitles, translation, meeting records, and interpreting all work for free. Live notes, one-step voice translation, and voice cloning require your own Ling (Ant Group's LLM service) API key. An hour of meetings costs roughly ¥0.15 to ¥1 RMB, and Ling provides ¥10 of free credit every month, enough for 10 to 50 hours.
Intel Macs, Windows, Linux, Android: Speech recognition runs in the cloud and requires an AI service key; audio is sent in segments to Ling's Ming or your chosen service (OpenRouter, ZenMux, etc.). Recordings and meeting records still stay only on your device.
The Mac app only asks for a permission when you use the corresponding feature, and everything else works normally if you decline:
| Permission | What it's for | Without it |
|---|---|---|
| System audio recording (Screen & System Audio Recording) | Hears system audio for subtitles; never records the screen | Subtitles and notes won't work |
| Microphone | Hold to talk or Hands-free for interpreting; click Turn on mic to listen to a meeting room | Can't speak to translate or listen to the room |
| Input Monitoring | Detects the right ⌥ key (doesn't block or log other keys) |
Switch to a ⌃/⌘ shortcut to skip this permission |
| Administrator password | Installs the BlackHole 2ch virtual microphone (asked once) | The other side can't hear interpreting; everything else works |
- Transparent networking: Only connects to the internet for translation (
translate.googleapis.com), your configured AI service (such asapi.ant-ling.com), downloading the Apple-notarized BlackHole driver (verified against SHA-256), and checking GitHub once a day for updates. No analytics, no tracking, and no server of our own. - Local storage: Records and recordings live in
~/Library/Application Support/LiveSubs. Your key and voice sample are readable only by your current system account and excluded from Time Machine and logs. - Hidden from screen capture: On Mac, screen sharing, screen recordings, and screenshots cannot capture the notes window. You can manually turn on Show in screen sharing when recording a demo.
The table below shows the setup on Apple silicon Macs. When using a self-hosted service (Other (OpenAI-compatible)), you can turn on Send audio and text only to your service in Settings › AI Service to bypass Google completely.
| Feature | With a Ling key | Without a key / fallback |
|---|---|---|
| Speech recognition and language detection | Apple SpeechAnalyzer / SpeechTranscriber (on device); macOS 14.4–15 uses SFSpeechRecognizer (on device, needs Dictation) |
Same |
| Translating finished sentences | Ling Ling-3.1-flash (can be changed to Ling-3.0-flash in Settings › AI Service) |
Google Translate → Apple Translation (on device, macOS 26 only) |
| Translating the sentence still being spoken | Google Translate | Shows original text until replaced by the full-sentence translation |
| Live notes and post-meeting full version | Ling Ling-3.1-flash |
Not available |
| Checking if the key works | Ling Ling-3.0-tiny (1 token) |
— |
| Interpreting: hearing and translating | Ming Ming-3.0-audio-flash (one step) |
Apple on-device recognition + translation chain above |
| Interpreting: voice output and voice cloning | Ming Ming-3.0-tts-0.5B |
Built-in Mac voices (AVSpeechSynthesizer), no cloning |
| Speaker identification | WeSpeaker ResNet34-LM (bundled, runs on device via Core ML) | Same |
- Sign up and log in on the Ling open platform, create a token, and copy the key starting with
sk-. - Open Settings › AI Service in Cicada (or the banner at the top of the notes window), paste the key, and save.
Can't sign up for Ling? At the top right of Settings › AI Service (or by clicking Can't use Ling? during onboarding), you can switch to OpenRouter, ZenMux, OpenAI, or any OpenAI-compatible service (including local or LAN services like Ollama, vLLM, SGLang, Qwen3-ASR, and Whisper).
After switching, live notes and full-sentence translation still work, and OpenRouter and ZenMux can also use natural Gemini voices for interpreting; one-step voice translation and voice cloning are supported only by Ling. With a self-hosted service, notes and speaker naming can use a stronger model of their own, or connect to another service (enter its own address and key); search for models in the list, or type one in directly if it isn't found. When Windows, Linux, or Android switches to another service, listening to the other side also goes through the new service, so subtitles run 1–2 seconds behind Ling. If a connection fails, Settings tells you directly whether it's a key, model, or proxy issue (see the setup page for proxy details).
Installers are on the Releases page and the website, and you can verify SHA-256 checksums against checksums.txt in Releases. If you run into questions, reach out to the author @wquguru on X.
Requires macOS 14.4+ (Apple silicon and Intel). On Apple silicon with macOS 14.4–15, first go to System Settings › Keyboard to turn on Dictation and add the languages you want heard; macOS 26 uses the new on-device model directly without needing Dictation; Intel Macs use cloud-based Ming recognition and require a Ling key.
Download Cicada-<version>-macOS-AppleSilicon.dmg or Cicada-<version>-macOS-Intel.dmg and drag Cicada into Applications (verify with shasum -a 256 Cicada-*.dmg). The app is not notarized by Apple yet, so if the first launch is blocked, go to System Settings › Privacy & Security and click Open Anyway. Follow the onboarding guide to grant permissions and try the sample meeting.
Requires Windows 10 2004+ (x64 or Snapdragon ARM64). Download the installer and double-click to install without admin rights. If SmartScreen blocks it, click More info › Run anyway. When you use interpreting, the guide walks you through installing the free VB-Cable virtual microphone.
Requires Ubuntu 22.04 or 24.04 desktop (x86_64 or arm64). Download Cicada-<version>-Linux.tar.gz and run tar xzf Cicada-<version>-Linux.tar.gz && cicada-linux/install.sh in a terminal. Interpreting automatically creates a "Cicada Microphone" with no driver to install. Linux windows cannot hide from screen sharing, so the notes window indicates that it is visible when sharing your screen. To enable Settings › General › Save meeting audio, first install system third-party media formats (such as ubuntu-restricted-extras) to provide an AAC encoder. Turning on Settings › Word list › Fix misheard words live sends one extra text request per sentence.
Requires Android 10+. Download and install Cicada-<version>-Android.apk; on Android 15 and later, allow restricted settings in App info the first time you turn on floating subtitles. It also includes on-device voiceprints to separate speakers and fixes misheard words live while listening. Call and meeting apps block internal audio capture on Android—just put the call on speaker and switch to In person.
In rough order of priority; feel free to open an Issue for what you'd like to see. See docs/ROADMAP.md for details.
- Record my own speech: Continuously listen to the microphone with echo cancellation during meetings so what you say directly is also included in records and notes.
- Code signing and notarization: Official signing and notarization for Mac and Windows installers to eliminate security prompts and move AI keys back to the system keychain.
- Speaker recognition: Remember frequent coworkers' voices on device and recognize them automatically in future meetings.
- Notes without a key: Generate meeting notes on supported Macs using Apple's on-device models.
- Android on app stores: No more manual APK installs.
- iPhone and iPad: Starting with listening to meeting rooms, face-to-face interpreting, and meeting records.
Quit Cicada and drag it from the Applications folder to the Trash. Meeting records and recordings are in ~/Library/Application Support/LiveSubs and can be deleted manually if no longer needed. To clear settings, run:
defaults delete com.wquguru.cicadaThe Mac app only requires Xcode Command Line Tools (xcode-select --install) and has no third-party dependencies. On macOS 27, if SwiftUIMacros is reported missing, install full Xcode and run sudo xcode-select -s /Applications/Xcode.app. For building on other platforms, see windows/README.md, linux/README.md, and android/README.md.
scripts/build.sh # dev build → build/Cicada.app ("Cicada Dev", runs alongside the release app)
open build/Cicada.app
scripts/package.sh # package release DMG, zip, and SHA-256 checksumsSee AGENTS.md for code architecture. The website source is in site/ (generated by scripts/site.py), and release history is in CHANGELOG.md or on the website changelog.
This project is licensed under Apache-2.0. The BlackHole 2ch virtual microphone driver is licensed under the GPL; it is not bundled in the installer and is downloaded from its official URL when you click install. The bundled voiceprint model is WeSpeaker's ResNet34-LM trained on VoxCeleb (CC BY 4.0), converted to Core ML by scripts/convert-voiceprint.py.