Lightweight Single-Channel Speech Enhancement Based on Multi-Frame MVDR Filters with Learnable Parameters
Haoyang Li, Zhanheng Zheng · 2024
In daily life, noise may interfere with speech signals, thus seriously affecting speech quality and intelligibility. To solve this problem, many traditional methods based on statistical modeling and methods based on machine learning have been proposed. Nevertheless, most of these algorithms inevitably suffer from a drawback: the attenuation of the noise component may lead to some distortion of the speech component in the enhanced signal. In this paper, we estimate the noise covariance matrix and speech interframe correlation vectors by employing DNN networks, and use them to construct a learnable parametric multi-frame MVDR filter as a way to fully utilize the short-term correlation within the speech signal and apply it to single-channel speech enhancement; we also evaluate the method with baseline model, and objectively demonstrate it's excellent performance and the less number of model parameters and computation compared to mainstream models.