Speech denoising is a prominent and widely utilized task, appearing in many common use-cases. Although there are very powerful published machine learning methods, most of those are too complex for deployment in everyday and/or low resources computational environments, like hand-held devices, smart glasses, hearing aids, automotive platforms, etc. Knowledge distillation (KD) is a prominent way for alleviating this complexity mismatch, by transferring the learned knowledge from a pre-trained complex model, the teacher, to another less complex one, the student. KD is implemented by using minimization criteria (e.g. loss functions) between learned information of the teacher and the corresponding one from the student. Existing KD methods for speech denoising hamper the KD by bounding the learning of the student to the distribution learned by the teacher. Our work focuses on a method that tries to alleviate this issue, by exploiting properties of the cosine similarity used as the KD loss function. We use a publicly available dataset, a typical architecture for speech denoising (e.g. UNet) that is tuned for low resources environments and conduct repeated experiments with different architectural variations between the teacher and the student, reporting mean and standard deviation of metrics of our method and another, state-of-the-art method that is used as a baseline. Our results show that with our method we can make smaller speech denoising models, capable to be deployed into small devices/embedded systems, to perform better compared to when typically trained and when using other KD methods.
