Speech Bandwidth Extension Using Generative Adversarial Networks

Speech blind bandwidth extension technologies have been available for some time, but until now have not seen widespread deployment, partly because the added bandwidth has been accompanied by added artifacts. In this paper, we present three generations of blind bandwidth extension technologies, from Vector Quantization mapping through Gaussian Mixture Models, to our latest architecture based on deep neural networks using Generative Adversarial Networks. This latest approach shows a sharp jump in quality, and demonstrates that machine-learning based blind bandwidth extension algorithms can achieve quality equal to wideband codecs, both objectively and subjectively. We believe that blind bandwidth extension can now achieve sufficiently high quality to warrant deployment in the existing telecommunication networks.