wquguru/cicada

What it does

Cicada is a macOS menu bar app that listens to the audio your Mac is already playing and produces live bilingual subtitles, on-the-fly structured meeting notes, and optional simultaneous interpreting. It works with any call or stream — Zoom, Feishu, Tencent Meeting, X Spaces, YouTube — because it captures system audio rather than joining the meeting as a bot. Speech recognition, language identification and recording run locally with Apple's SpeechAnalyzer/SpeechTranscriber; recognition text is sent out only for translation and note generation.

Meeting records keep the bilingual transcript, notes and raw recording in ~/Library/Application Support/LiveSubs, with Markdown, Word and SRT export. Optional speaker diarization uses a bundled WeSpeaker ResNet34-LM Core ML model.

Getting started

1. Check the machine: macOS 26 or newer on Apple silicon.
2. Download Cicada-<version>-macOS-AppleSilicon.dmg from the Releases page or the official site, verify it with shasum -a 256 Cicada-*.dmg against the published checksums.txt, then drag the app to Applications.
3. The build is locally signed but not notarized, so on first launch open System Settings → Privacy & Security and choose to open it anyway.
4. Follow onboarding: grant screen/system audio recording, play the built-in two-person sample meeting to see subtitles, and optionally paste a Ling API key.
5. For interpreting, install the BlackHole 2ch virtual microphone when prompted; hold the right ⌥ key to speak, and the other side hears the translated voice.

The app requires no third-party build dependencies. From source: install Xcode Command Line Tools, then run scripts/build.sh and open build/Cicada.app.

Using Ling and network access

Live notes, one-step speech translation and voice cloning need your own Ling API key, created on the Ling open platform and pasted into Settings › AI services. The README estimates roughly ¥0.15–1 per hour of meeting and notes that Ling grants ¥10 monthly credit, about 10–50 hours. Without a key, subtitles, translation (Google, falling back to on-device Apple Translation), meeting records and interpreting with the built-in Mac voice still work; notes do not.

Only translation, Ling/Ming endpoints, the BlackHole download and a daily GitHub update check go online. There is no telemetry or self-hosted server. The README documents each permission — system audio recording, microphone, input monitoring for the right ⌥ key, and admin password for installing BlackHole — and states that declining one leaves the other features usable.

Constraints and roadmap

• macOS 26+, Apple silicon only.
• The current build is not notarized; first launch needs the Privacy & Security override.
• The transcript currently records your own speech only when you hold the talk key, not what you say directly in the call.
• Planned work includes recording your own speech, support for macOS 14.4 and later, notarization, voiceprint memory, on-device notes without a key, and Windows/Android; iPhone/iPad are not planned due to platform limits.

HotRepo's take

The project appeared on HotRepo's new-project list on 2026-11-03, gaining 73 stars and 18 forks in the recorded one-hour window (170 → 243 stars) with no GitHub all-language daily rank recorded. It stands out less for raw size than for the concrete macOS-native approach: real system-audio capture, an explicit permission/privacy table, and a documented free path. The open question a reader should weigh is the un-notarized distribution plus the reliance on Ling for the signature notes feature — both are acknowledged in the README's roadmap rather than hidden.