Audio Denoising using Encoder-Decoder Deep Neural Network in the Case of HF Radio

This paper presents an artificial intelligence approach to speech enhancement, demonstrated through the restoration of highly damaged and noisy HF radio demodulated signals. Noise is an omnipresent phenomenon that significantly degrades the listening experience. To mitigate its effects, we introduce a method based on the encoder-decoder deep neural network architecture. Encoder-decoder neural networks, renowned for their versatility and capacity to learn meaningful representations and generate data based on those representations, have found applications across various domains such as dimensionality reduction, feature learning, data denoising, image generation, anomaly detection, and language modeling. When compared to different neural network architectures, this approach outperforms others in objective speech quality metrics and perceptual experiments with volunteer listeners. Furthermore, it surpasses traditional speech denoising algorithms. The advantage of this method is particularly evident for highly distorted signals undergoing severe speech degradation, where denoising is both most needed and most challenging.