Image: GitHub
Why it mattersA buyer who was paying for Otter or Granola can keep the audio on one machine. The app stays usable with no API key.
Most meeting transcription apps join the call as a bot. Everyone on the call sees the bot name in the attendee list. The audio is then streamed to a vendor's server for transcription. Cicada is a Mac app that does the job without joining the call. The author @wquguru put it on GitHub on 1 November 2026. It reads whatever is playing through the Mac's system audio. It writes bilingual subtitles, a running transcript, and a structured summary.
The repository is wquguru/cicada. It had 545 stars and 117 forks at the time of writing. It ranked 14th on Trendshift's daily list. The licence is Apache 2.0. The app is Swift and SwiftUI. It runs on Apple Silicon with macOS 26 or newer. The installer is a signed but not yet notarised DMG from GitHub releases.
What it does during a call
The README says Cicada captures system audio through macOS's screen recording API. The audio is fed to Apple's on-device SpeechAnalyzer model for recognition. Half-sentences are translated live through Google Translate as the speaker talks. The Google call does not count against the paid API key's rate limit. Completed sentences are translated by Ant Group's Ling 3.1-flash model if a key is set. Without a key, Apple's on-device Translation framework is used. The subtitle bar appears only while someone is speaking. It goes transparent when the mouse moves over it.
A simultaneous interpretation mode lets the user hold the right Option key to speak. The app replays the translation through a virtual microphone named BlackHole 2ch. The other side then hears the translated voice in Zoom, Feishu or Tencent Meeting. With a Ling API key, the voice can be cloned from an 8 to 30 second recording. Without one, macOS's own speech synthesiser is used.
What it does after the call
The structured notes have sections for topics, conclusions, action items, a fishbone diagram, and a glossary. The glossary unifies spellings of names and terms. Clicking any line in the summary jumps to the matching bilingual transcript line. Transcripts, audio and notes are stored locally under ~/Library/Application Support/LiveSubs. Exports are Markdown, Word, or SRT subtitles. An experimental speaker identification mode uses the open WeSpeaker ResNet34-LM voiceprint model. It is included with the app and runs through Core ML.
What it costs
The author says a one-hour meeting costs 0.15 to 1 yuan on Ant Group's Ling platform. Ling gives a free 10 yuan monthly allowance. That covers about 10 to 50 hours of meetings a month. Without a key, the app still runs live subtitles, translation, meeting recording, and simultaneous interpretation. The structured notes and the one-step speech translation are the only features that need the key.
Source
cicada on GitHub, repository by @wquguru.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
