Code & Sundry

A Tiny Test of Krisp.ai

314 words, 2 minutes to read

And I mean tiny!

If you haven't heard about Krisp.ai before, it's a pretty cool concept. It's software noise cancellation, built using a deep neural network. They've built an algorithm to filter out background noises by analysing thousands of hours of voice recordings from a lot of different people, which will recognize your voice and filter out any noise from your surroundings.

I also like (or actually it's more like love) that they have a focus on privacy. All audio processing is done on-device, not through some cloud service. BIG THUMBS UP!

I was curious though, how good is it really?

I downloaded the app which, it turns out, allows you to make a 15 seconds test clip to see how well it works. While they do delete all sound clips as soon as they're done with them, they need to store them somewhere while they're being used by the software. I located the files on disk and stored for your sampling pleasure.

The original

With noise cancellation

Without noise cancellation

If this is representative of the quality you can expect, I'm not too impressed. Comparing the original and the clip without noise cancellation, the difference is... substantial.

Comparing the clips with and without noise cancellation though, is quite interesting. While they are somewhat poor quality, it's cool how it's able remove to most of the noises I'm making.

My assumption is that they're going to work on improving the algorithm and it'll probably (hopefully?) get better with time. For now though, I'm not onboard. I'd rather have the superior audio quality, than the noise cancelled one. But I do have the advantage of a good microphone and a quiet place to use it.