Krisp got its start as an AI-powered noise-cancellation service, but now the startup is moving beyond that with instant on-device transcription of all your calls and meetings — and of course it integrates ChatGPT now for quick summaries as well. The open beta goes live today.

My transcribing experience is usually a matter of downloading the video or audio associated with a meeting, provided someone thought to record it (and I know who that someone was), then uploading that file to a service like Otter and waiting 10-15 minutes while it processes. Compared to just a few years ago this is very convenient, of course, but the number of meetings we take online has multiplied. Wherever I can shave a few minutes or clicks off my workflow I'm usually interested.

Krisp has also shown a certain single-mindedness over the years, focusing on effective noise reduction across lots of devices and services. The advance that made it possible was basically training the system to recognize and isolate human voices, which while diverse are also distinct and very different from dogs barking, traffic noise and so on. For me, Krisp has generally Just Worked.

So when they told me they were getting into transcription I was happy to hear it: It's a natural extension of a voice-focused platform, and of course they know they can get a clean voice signal in.

Krisp already operates more or less behind the scenes, acting as a passthrough for audio sources. So there's no need for the "AI Meeting Assistant" to integrate with Google, Teams, Zoom or anything else, or be a phantom attendee — it deals strictly with your audio signal before it goes out or before theirs hits your ears. The transcription agent doesn't need to talk to any services or APIs, it just identifies voices coming in and going out, and transcribes them, on your device.

On-device transcription is a key differentiator, for one thing because it's simpler than dealing with audio files, and for another because the audio of your meeting never leaves your computer. I'm not sure what threat models that's important to, but I suspect they exist and will only get more prominent in the coming years.