Krisp goes from noise cancellation to on-device transcription

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

My experience with transcription is usually to upload the video or audio associated with a meeting, provided someone remembered to record it (and I know who it was), then upload that file to a service like Otter and wait 10 to 15 minutes while processing it. Compared to a few years ago, this is of course very convenient, but the number of meetings we organize online has multiplied. Anywhere I can save a few minutes or click through my workflow, I’m usually interested.

Krisp has also shown some determination over the years, focusing on effective noise reduction across many devices and services. The breakthrough that made this possible was basically training the system to recognize and isolate human voices, which, while diverse, are also distinct and very different from barking dogs, traffic noise, and the like. 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-centric platform, and of course they know that they can get a clear voice signal.

Krisp is already operating more or less behind the scenes, acting as a relay for audio sources. So there’s no need for “AI meeting assistant” to integrate with Google, Teams, Zoom or anything else, or be a ghost participant – it strictly processes your audio signal before it goes out or before theirs reaches your ears. The Transcription Agent doesn’t need to talk to any services or APIs, it just identifies incoming and outgoing voices and transcribes them to your device.

On-device transcription is a key differentiator, partly because it’s easier than processing audio files, and partly because your meeting audio never leaves your computer. I don’t know for which threat models this is important, but I suspect they exist and will only grow in importance in the years to come.

The transcription itself is sent directly to Krisp’s cloud service, which kind of negates the on-device security benefit of the transcription. But if you’re ready to trust another transcriptionist, why not this one? And anyway, there will soon also be an option to keep your transcriptions completely on the device. Krisp CEO and co-founder Davit Baghdasaryan has a strong background in security and said he was eager to implement this. (I suspect it’s just easier to build a web app than to add major new features to multiple mobile and desktop apps.)

And how are the transcripts? In my experience about as good as the other AI powered ones i.e. better than my ratings but not quite ready to release. Dropping the audio immediately defeats a useful feature offered by other transcribers, quick checking and editing: click on a suspicious word and it jumps to that place in the audio. So, for now at least, in interviews where you need to quote accurately and in detail, it might not be the best choice.

But Krisp positions it more as a productivity tool, creating a platform to quickly extract key points from the countless meetings we sometimes find ourselves scheduling. To this end, the company allows you to quickly send the transcript for analysis by ChatGPT, which extracts the TL; DR, assigned actions, or whatever needs to be pulled from the meeting.

Frankly, I never liked these features and never use them, so I can’t speak to quality here. They are probably as good as any other service that does this, i.e. helpful to people who find such things useful. I try to forget the content of a meeting as quickly as possible, no doubt to the chagrin of my colleagues, but in my defense, I would remember it badly. Organized people who have processes for this type of work may or may not find the Krisp style to suit them.

I’m also hesitant to judge a service that is essentially an API call to another service, which itself is constantly evolving. As ChatGPT improves, so will the apps that depend on it – like here, like on a hundred other platforms that have suddenly developed an ability to understand natural language.

Krisp is free if you want to try it out – the company makes money on premium subscriptions for business, and of course, with the compute running on your device, it’s pretty simple to scale.

It’s worth taking a moment on your Mac or PC to make sure it’s set up correctly, i.e. its relay is working as expected. I had some weird issues at first, but I fixed them by starting from scratch and only assigning meeting apps and mics to go through Krisp, so it wouldn’t mix up my music. It’s just a handful of dials and switches, so take five minutes to get it right and soon your transcripts will appear so fast it’ll be a bit boring.