Enhanced MVDR beamforming for Arrays of Directional Microphones

Опубликовано 7 сентября 2016, 16:24

In many signal processing tasks, it is of interest to estimate some desired signal in the presence of noise as well as other sources of interference. If the desired and interfering signals have overlapping spectral content, estimation of the former from the observed mixture cannot be successfully achieved by so-called SISO (single input single output) filtering methods. More often than not, the desired signal and the sources of interference are emitted from distinct spatial locations. This has motivated the use of MISO (multiple input single output) systems, also known as arrays, which allow for more degrees of freedom in the design of an effective filter. Beamforming (or spatial filtering) refers to the use of arrays of sensors to discriminate between a set of signals based on the spatial location from which they emanate. The well-known Capon or MVDR (Minimum Variance Distortionless Response) beamformer is one of the most pervasive ways of performing spatial filtering. Assuming the DOA of the desired signal is known, the MVDR beamformer is able to estimate the desired signal while minimizing the variance of the noise component of the formed estimate. In practice; however, the DOA (direction of arrival) of the desired signal is not known exactly, which significantly degrades the performance of the MVDR beamformer. Even assuming 100 SSL (Sound Source Localization) accuracy, the fact that the sensors may have distinct, directional responses adds yet another level of uncertainty that the MVDR beamformer is not able to handle well. In this talk, we propose an improved MVDR beamformer which takes into account the effect of sensors with arbitrary, potentially directional responses. Specifically, we form estimates of the magnitude responses of the sensors based on the data received at the array and include those in the original formulation of the MVDR beamforming problem. Preliminary results of experiments on real data show, on average, a 2.6 dB improvement over conventional MVDR beamforming (which does not account for sensor response) and a 3.6 dB improvement over a naïve scheme which selects the sensor with the highest SNR as the estimate of the desired signal. We are currently conducting a Mean Opinion Score (MOS) test, the results of which shall be available in the near future.