Speech enhance methods base on traditional digital signal processing (DSP) algorithms or adaptive filters can effectively suppress stationary noises. However, they don't provide viable solution for the variety of non-stationary noises that exist in our everyday life. Smart voice assistants such as Google Home and Alexa deteriorate their performance mostly due to non-stationary noises. In this paper we introduce CycleGAN ANF, a neural network approach that can learn to reduce both stationary and non-stationary noises, totally unsupervised. CycleGAN ANF is capable of reducing undesired interference by reading in a raw audio sample from a set X (speech mixed with noises) and transforming it so that it sound as if it belongs in set Y (clean speech). Our experiments demonstrate that without labels and when trained on unparalleled; relatively small vocabulary of speech datasets, CycleGAN ANF can achieve significant improvements without the ground assumptions of nature and form of the noise.