Final Results

Now that we've seen two ways of improving the quality of our signal, it's time to combine them. There are two ways to do this. The more efficient way is place a single echo canceller after the microphone array. The problem with this method is the beamformer is now a part of the room echo response. And because the beamformer can change its look direction rapidly, the echo canceller will not be able to track the response fast enough [BW]. This means the echo canceller will not work when the talker moves.

The other method is to place an echo canceller on each microphone. This makes the amount of computation depend on the number of microphones. There are some things we can do to decrease the workload. One is to use a block exact algorithm. Also, some of the work (inverting the correlation matrix) done by the FAP algorithm depends only on the reference signal. So, this computation is done only once. Beyond these few tricks, we'll just have to optimize our code and find a Digital Signal Proecessor (DSP) that can handle the work.

Here is the output of a 16 microphone simulation. Some of the noise can be removed from the signal using a thresholding technique. If you compare the second signal to the first one, you might notice it sounds like the talker has a hand over his mouth. This is because the delay-and-sum beamformer acts as a low pass filter. There are ways of dealing with this. By applying a Finite Impulse Response (FIR) approximation of an inverse filter, there is some improvement in the resultant signal. For this simulation the time delays were known ahead of time.