← Back to blog
AiAbout 6 min read

Cicada listens to your meeting's audio directly, with no bot in the room

Published Nov 21, 2026
Cicada listens to your meeting's audio directly, with no bot in the room

The first pleasant surprise about Cicada is what it does not do. It does not join your meeting. There is no participant named "Cicada Notetaker" waiting in the lobby, no calendar permission, no email to the other side explaining that an AI is listening.

It captures the audio your computer is already playing. That single choice removes most of the friction that meeting bots create, and it also removes their excuse: nothing in the call changes, so nobody has to agree to anything.

Cicada, by wquguru, is free, Apache-2.0 licensed, and appeared on GitHub on 1 November 2026. It had 2,437 stars three weeks later.

What is on screen

A subtitle bar sits at the bottom, shows up when somebody speaks, turns see-through under your pointer, and translates into your system language after detecting the spoken language within a few seconds. ⌃⌥S cycles display modes if you want source and translation together or separately.

On the right, notes build as the meeting runs: a summary, topics, key points, conclusions, to-dos, a fishbone diagram, and a glossary of names and terms. Click a conclusion and it jumps back to the lines it came from. ⌘F searches the transcript, and homophones that were misheard still turn up, because the search runs over corrected text rather than raw audio.

The correction loop is the part I did not expect. With an AI service connected, misheard names get fixed on the spot: obvious homophones immediately, unfamiliar names by asking you once and remembering your answer. You can paste a word list into settings before a meeting so the vocabulary is there from the first minute. After the meeting, Edit line… batch-fixes the same wrong word across the whole session, with undo.

Two kinds of audio, one record

Computer audio is labelled Online. The microphone is labelled In room, and turning it on captures the people physically with you. Both go into the same record, which is the correct behaviour for a hybrid meeting and surprisingly rare.

With speakers instead of headphones, Cicada deduplicates the audio so the same voice does not appear twice. With headphones, both sides can talk at once and both are kept. On a Mac, the arrow next to the mic button picks a specific input, an iPhone mic or a conference mic, without changing the system default.

What the microphone hears only goes into the notes; it is not routed into the meeting. For the other side to hear you, you use the interpreter.

Hold a key, speak your own language

This is the feature that made me look twice. Hold the right ⌥ on a Mac, or the right Ctrl on Windows, and speak. The other participants hear your translated voice inside the call, in a voice that can be yours: with a Ling API key, Cicada clones it from an 8 to 30 second recording. Without a key, you still get the built-in macOS voices.

On a Mac there is also a hands-free mode: start talking while the meeting plays, the meeting audio ducks automatically, and your translated voice goes out after a one second pause. The mic pauses while it is on, and the README recommends switching back to the talk key when other people are nearby so their voices do not get mistaken for yours. Practical advice, and easy to forget.

Under the hood, turning on interpreting switches the system microphone to BlackHole 2ch, a virtual audio driver you install with one click. This is the one setup step that will confuse people, and it is unavoidable: to inject audio into a meeting app, something has to pretend to be a microphone.

Speakers, names, and the record

The on-device voiceprint model separates speakers automatically, and fills in names when a Ling key is present. You can rename or merge speakers whenever it gets one wrong, and on a Mac you can tell it to remember a voice, with the README's own instruction to check with that person first. The voiceprint stays on the machine, and settings let you delete it.

Records keep the bilingual transcript, the notes and the raw audio on your device. Click any line to play from that point while the text highlights. After the meeting you can regenerate notes, edit a conclusion in place, re-transcribe if the language was detected wrong, or import a local audio or video file. Export covers Markdown, Word and SRT.

Two details stand out. Copy Notes pastes into Feishu, Notion or Word with headings, lists and the fishbone picture intact. And with "Save to folder" on, every meeting becomes a Markdown file with frontmatter at the top, dropped into an Obsidian vault or any folder, updating when you rename it. Frontmatter for agents, in a consumer app, in 2026: the assumption is that something downstream will read these notes.

Sharing generates a link with optional expiry and password, and before uploading it blocks keys, unpublished company information and other sensitive strings. Blocking is a filter. It is the right default direction anyway.

Platform reality

Mac has the full feature set. On Apple silicon, speech recognition runs on device, so audio never leaves the computer and you can use the app without an API key. Intel Macs need a key. Windows, Linux and Android need a key from Ling, OpenRouter, ZenMux or similar.

On Linux the notes window is visible in screen sharing, which you should know before presenting. On Android, the system blocks recording audio from call and meeting apps, so the workaround is speakerphone plus the In person mode. That is a platform limit, not a bug.

What I would check before trusting it

Four things.

Consent. "No one else notices a thing" is a feature for the person recording and a problem for everyone else. In many jurisdictions recording a call requires informing participants. The app makes it easy to forget that; you should not.

Translation quality under pressure. Live interpreting is a hard task, and a one second pause before your voice goes out is a real constraint on how you speak. Try it in a low-stakes call before you use it with a client.

The dependency chain. On-device recognition is Mac-only, and everything else needs a third-party key. Free software with a paid dependency in the middle is a normal arrangement, and it is worth knowing which parts stop working when the key expires.

Speaker attribution. Voiceprints are good and not perfect, especially with similar voices on a compressed meeting codec. Read the first minutes of a transcript before you trust the labels.

Why it is worth the disk space

There are three separate products here that usually get sold separately: live subtitles, meeting notes, and real-time interpreting. Cicada ships all three, uses the audio the computer already has, and stores the results as files you own.

The design decision that makes it small is the absence of a bot. Everything else follows: no calendar integration, no lobby, no participant list, no "let me in" step. You press a key, and the meeting gets subtitles.

For anyone who spends their week in calls across two languages, that is a large amount of friction removed for the price of one virtual audio driver.

Related articles