It’s been a while since my last post. I took a break from stem separation and focused on speech dereverberation (and noise filtering).

Just like stem separation, this is something that would have been impossible before AI.

So it’s something that really excites me. I decided to try to make something that makes a distant recording sound like the speaker was sitting in front of a microphone.

I’m excited to say that it has good results. I tested it by recording a conversation with another colleague of mine with my cellphone on a table. And it worked pretty well.

The difference between most models of noise reduction and mine is that this was not designed to work on realtime, so it works on chunks of 1.1 seconds. The model is not complex in its architecture in terms of ops, it uses 1D/2D convolutions with LeakyReLu activations and transformer layers, basically a standard U-net in time and frequency domains. Self-attention at the bottleneck, which helps to get better results. 

The model adds a subtle distortion to the audio in some cases, which probably has to do both with the upsampling method and the loss function. But still sounds pretty good on my test data.

So, I invite you to check the official site I made for this at:

http://dereverb-ai.gfsoftware.com

Also… I definitely plan to take this new model that takes advantage of time and frequency domain analysis and see if it improves the stem separation models I already have.

Have a listen !

Audio demo

Original

Processed