Kalypta, that's what it's called, protects conversations from covert transcription. Right now a ton of modern call-recording and transcription apps don't need a bot or agent added to the call — they just capture everything straight off your computer: Granola, Whisper, Notion, MyMeet, and others. Privacy takes a hit.
How it works
The app alters your outgoing audio signal locally, in real time. The person you're talking to hears you just fine, while automatic speech recognition systems get garbled text.
Kalypta adds adversarial noise to the audio stream — deliberately calculated distortions designed to throw off speech recognition models. Meanwhile, the sound a human actually hears barely changes.
It's basically the audio version of an idea from my post about fonts that humans read fine but computer vision systems can't.
Results and limitations
According to the developers, Kalypta on average blocks recognition of roughly two words out of three. The model runs locally on-device and was trained against the architectures behind Whisper, NVIDIA Canary, and other transcription systems.
There's an important caveat (of course there is): the protection isn't perfect yet. Crank the distortion up enough and speech gets harder for humans to follow too. Also, the published results are the team's own early tests, not an independent benchmark.
But it's a fun direction. Kind of reminds me of early Krisp, back when the whole project was just about stripping background noise out of calls — that was the entire product (now they're a unicorn, and they record calls).