A Joint Bandwidth Expansion and Speech Enhancement Approach Using Deep Neural Network

Abstract

Recently, joint bandwidth expansion and speech enhancement has been a topic of interest in the field of speech processing. The main challenge in this task is to increase the bandwidth of speech signals while enhancing their quality, simultaneously. Deep neural networks (DNNs) have shown great promise in addressing this challenge, as they can learn complex relationships between the input and output signals. In this study, a joint bandwidth expansion and speech enhancement approach using DNNs have been proposed, which is designed to simultaneously increase the bandwidth of speech signals and reduce noise, while preserving speech quality and intelligibility. This approach leverages the capability of DNNs to simultaneously estimate the missing speech components and the noise profile in the degraded speech signal. The estimated speech components and the noise profile are then used to synthesize a full-band speech signal from a noisy signal with limited bandwidth with improved quality. The network employs three different phases such as oracle, imaged, and noisy phase along with the magnitude spectra to recover high band components. The joint approach demonstrates that the DNN-based bandwidth extension and speech enhancement can be effectively combined to produce high-quality speech signals, outperforms traditional speech enhancement methods, and offers promising solutions for various applications, including speech communication, speech recognition, and speech synthesis.