 Hello, everyone. My name is Artigral, and I'd like to shortly present to you our approach for noise suppression that uses features of the sound textures, which are class of sounds that are usually encountered as noise in our normal acoustic environments. So what are the sound textures? These are a class of sounds that are composed of simpler acoustic events that occur repetitively and are not very variable in terms of their spectro-temporal structure. Think of the sound of the stream, a train, or a flock of geese. This category of sounds became an active field of research after the work done by McDermott and Simon Shelley that showed that realistic textures can be created from white noise using a small set of statistical features. This model now allows us to control naturalistic noises experimentally, which is something that wasn't possible before. Our approach leverages sound textures in two main ways. First, we use the limited variability of the acoustic components to train our algorithm. And second, we evaluate general effectiveness of our method on various background noises that span a wide range of values in statistical space instead of solely relying on pre-recorded clips. During the training stage, we create a library of spectro-temporal features of the background noise using a k-dimensional tree representation. We obtain this training data from the user in a supervised mode or extracted it automatically using voice activity detection. Then we find the best matching noise fragments and subtracted it from the original mixture to yield clean signal. We then assess our method using metrics such as correlation to the original clean speech or other automatic ways of evaluating the sound quality in addition to running online experiments. And all of these metrics show that we are able to remove the textural noise effectively. Now the filtering methods that are to be used in devices like hearing aids need to be fast and computationally cheap. To see how our algorithm fares in this respect, we varied its parameters and analyzed the processing times. And our findings indicate that parameters required for optimal performance strung with black dots here allows our algorithm to run significantly faster than real time, which opens up the possibility to integrate it to existing processing pipelines. So in conclusion, we think paying attention to and explain features of what we commonly call noise is an effective avenue when it comes to developing noise suppression algorithms. And we propose a concrete method that achieves just that. Thank you for your attention. And don't forget to tune into our presentation if you want to learn more about it.