Improving deep speech denoising by Noisy2Noisy signal mapping

Existing deep learning-based speech denoising approaches require clean signals to be available for training. This paper presents a approach improve in real-world audio environments by not requiring the availability of as reference training mode. A fully convolutional neural network is trained using two noisy realizations same signal, one used input and other target network. Two signal are generated mid-side stereo microphone. Extensive experimentations conducted show superiority developed over conventional supervised based on four commonly performance metrics well subjective testing.