Eigen Beamforming

Download as pdf or txt
Download as pdf or txt
You are on page 1of 1

2.6.

EIGENBEAMFORMING

Eigenbeamforming consists of two steps: decomposition and synthesis [14]. The


decomposition step transforms the multichannel sensor signals into an orthonormal
space. Each base vector is called an eigenbeam. It is so named based on the analogy
of an eigenvector to a matrix. The synthesis step forms the desired array response
by weighting the eigenbeams and summing the results together. Theoretically, any
directivity pattern that an eigenbeamformer can attain can also be realized using
the additive beamforming method; but eigenbeamforming offers some unique advantages from an engineering viewpoint [15]: 1) it requires fewer signals to be stored
since the number of eigenbeams can be much less than the number of sensors; and 2)
it can form a desired response in a computationally very efficient way. However, to
use eigenbeamforming, the array has to be carefully designed so that the sensor positions can meet the orthonomality condition. This is a nontrivial job. Moreover, the
number of sensors with this beamformer is usually large to guarantee the required
spatial resolution. That is why this technique is currently used mainly for spherical
arrays even though theoretically plausible also for other shapes such as cylindrical,
oblate, and prolate. The magnitude response of the eigenbeams is dependent on the
eigenbeam order and has a highpass filtering nature. So frequency equalization is
required, as in differential beamforming.

2.7

NOISE REDUCTION

Noise reduction techniques intend to mitigate the effect of additive noise. This
noise can come from various sources and can profoundly affect the processing and
perception of speech signals in voice communication. Noise reduction is typically formulated as an estimation problem where the optimal estimate of the clean speech is
achieved by optimizing some criteria,such as the mean-squared error (MSE) between
the clean speech and its estimate, the signal-to-noise ratio (SNR), the a posteriori
probability of the clean speech given its noisy observations, etc.
The complexity of this problem depends on the number of accessible microphones.
Most of todays communication terminals are equipped with only one microphone.
Existing single-channel noise reduction techniques fall into one of the following three
classes [17]: filtering, spectral restoration, and model-based methods. The basic
principle underlying the filtering technique is to pass the noisy speech through a
filter/transformation. Since speech and noise normally have very different characteristics, the filter/transformation can be designed to significantly attenuate the
14

You might also like