Tensor completion algorithms for estimating missing values in multi-channel audio signals

Abstract

Audio inpainting is a widely used technology in the real world since audio signals with missing data are pervasive in many scenarios. The majority of existing works address the time gaps in single-channel audio signals, while completing multi-channel audio signals is rarely investigated.In this work, we tackle this issue using four different tensor completion algorithms and we evaluate them on speech audio datasets with gaps in the time domain. Based on extensive quantitative and qualitative experiments, the tensor completion algorithms generally achieve a superior predictive performance, including when the gap duration of the signals reaches values of up to 200 ms. Specifically, the experimental results illustrate that all of the applied tensor completion algorithms yield at least 56% improvement in signal restoration performance compared with single-channel based methods. Therefore, the tensor based approaches can capture the underlying latent structure over different channels to reconstruct incomplete multi-channel data.