A menu-bar interpreter that lives entirely on your Mac

What it does

Cicada is a native SwiftUI menu-bar app that listens to whatever audio your Mac is playing — Zoom, Feishu, YouTube, X Spaces — and produces live bilingual subtitles plus structured meeting notes (summaries, action items, conclusions, even a fishbone diagram) in real time. When it’s your turn to talk, you hold the right ⌥ key and speak; the other side hears you in their language through a virtual microphone. It’s a one-app answer to “I’m in a meeting in a language I half-speak.”

The interesting bit

The privacy architecture is the real design: audio capture, language detection, and speech recognition (Apple’s SpeechAnalyzer) all run locally, and no bot joins the call — the other participants never know you’re using it. Only recognized text leaves the machine for translation and summarization, and even that falls back to Google or on-device Apple translation if you don’t supply an API key. The notes window is even invisible to screen sharing and screenshots.

Key highlights

• Live bilingual subtitles with automatic language detection, appearing only while someone speaks
• Real-time structured minutes with clickable links back to the original transcript lines, plus fuzzy search that also finds misrecognized name/term variants
• Push-to-talk simultaneous interpreting with optional voice cloning from an 8–30 second recording (via Ling/Ming models)
• Experimental on-device speaker diarization using a bundled WeSpeaker ResNet34-LM Core ML model
• Records, audio, and notes stored locally; exports to Markdown, Word, and SRT
• No analytics, no vendor servers; permissions requested per feature and each has a documented fallback

Caveats

• Requires macOS 26+ on Apple Silicon — a genuinely high bar that excludes most Macs today.
• The app is locally signed but not notarized, so first launch requires the manual “open anyway” dance.
• Real-time minutes and voice cloning need your own Ling (Ant Group) API key; without it you get subtitles, translation, records, and interpreting, but no AI-generated minutes. Your own spoken contributions are only recorded when using push-to-talk, not when you talk normally.

Verdict

Worth a look if you attend multilingual meetings on a recent Apple-silicon Mac and care about audio staying local. Everyone else — Intel Macs, Windows, anyone on macOS 25 or older — is firmly in roadmap territory.