Multi-scale Generative Adversarial Network for Speech Bandwidth Extension

The purpose of speech bandwidth extension is to recover the high frequency information missing in narrowband speech signal. In recent years, with the introduction of neural networks, the research of speech bandwidth extension has been greatly developed. In this paper, we propose a new speech bandwidth extension method, which increases the sampling rate of speech signal from 8kHz to 16kHz and 48kHz respectively. The model architecture is based on a multi-scale generative adversarial network (MSGan) with multiple discriminators in both time and frequency domains, and is trained using time domain loss, frequency domain loss, adversarial loss and feature matching loss of the discriminators. The multi-scale waveform discriminators are able to improve the naturalness and clarity of speech. In addition, the combination of time domain and frequency domain can stabilize the training and balance the weights of different perceptual quality factors, thereby improving the perceptual quality of wideband speech. We compare our method with several baseline methods. The results show that our method outperforms other baseline methods on Log-Spectral Distance (LSD) and High-Frequency Log-Spectral Distance (H-FLSD), achieving a good balance between time-domain metrics and frequency-domain metrics.