Multi-objective noisy-based deep feature loss for speech enhancement
Rafał Pilarczyk, Władysław Skarbek · Photonics Applications in Astronomy, Communications, Industry, and High-Energy Physics Experiments 2019 · 2019
Deep neural networks have become a great tool for creating solutions to denoise the speech signal, improving the intelligibility, speech quality and signal-to-noise ratio. An important element during training deep speech networks is the use of an appropriate loss function that allows to improvement the subjective and objective measures. In our work, we used the loss function based on a well-trained deep network to classify whether the signal is noisy and clean. Thanks to this, the deep network responsible for denoising is based on minimizing the difference of deep features of the pure and enhanced signal. Our work shows that the use of only deep features in the loss function allows a significant improvement in the measurement of speech signal quality. Novelty is also feature extractor, which has been trained as a multi-objective noise classifier. We believe that deep-feature loss could help in the optimization of functions difficult to differentiate.