Transformer-Based Virtual Microphone Estimator

The performance of microphone array signal processing algorithms commonly depends on the number of microphones used, such as beamforming, blind source separation (BSS), and direction-of-arrival (DOA) estimation, to name but a few. Generally, a higher number of microphones often leads to higher performance. However, in real-world applications, with the constraint of the cost of the microphone, the number of microphones is always fixed, which may lead to a performance limitation of the algorithms in a more adversarial environment. In such cases, a technique that estimates the signals of the virtual microphone and mimic the real signals has been proposed called the virtual microphone (VM). To improve the estimation performance of the present VM method, in this paper, a time-domain Transformer-based method called Transformer-based virtual microphone estimator (TVME) has been proposed. By employing a fully supervised learning framework, the proposed TVME can be trained using the multi -channel observations from the real microphone recordings. The estimation performance of the proposed TVME is verified through a minimum variance distortionless response (MVDR) beamformer by utilizing the estimated virtual microphone signals as inputs. Simulation results indicate that the proposed TVME can estimate the signal of the virtual microphone more precisely compared with the baseline systems, which leads to a better enhancement performance of the beamformer.