Speech Enhancement with Cycle-Consistent Neural Networks

Deep neural networks (DNNs) have become a standard approach for solving problems of speech enhancement (SE). The training process of a neural network can be extended by using a second neural network, which learns to insert noise into a clean speech signal. Those two networks can be used in combination with each other to reconstruct clean and noisy speech samples. This thesis focuses on utilizing this technique, called cycle-consistency. Cycle-consistency improves the robustness of a network without modifying the speech-enhancing neural network, as it exposes the SE network to a much larger variety of noisy data. However, this method requires input-target training data pairs, which are not always available. We use generative adversarial networks (GANs) with cycle-consistency constraint to train the network using unpaired data. We perform a large number of experiments using both paired and unpaired training data. Our results have shown that adding cycle-consistency improves the models' performance significantly.

speech enhancement, GAN, generative adversarial networks, deep learning, cycle-consistency