COMPAS ID: 152113
The adaptive filter can dynamically identify echo and subtract it from the receive path. The
process of identifying the echo and determining how much echo to remove is called "training" or
“convergence.” The training process has certain requirements in order to be successful.
When the adaptive filter is fully trained it can eliminate most but not the entire echo in the receive
direction.
Looking at the "Echo cancellation: Success case" slide, the red triangles in the transmit direction
represent the signal going out to the network. Some of that may be reflected back in the receive
path as echo. (The few red triangles seen in the receive signal of most blue triangles.) The
adaptive filter looks at the transmit signal and compares it to the receive signal it gets from the
network. The circuit is able to identify possible echo and minimize it. The resulting signals are
then sent to the NLP. Echo is still present but is much reduced as evidenced by the much smaller
red triangles.
The signal is sent to the NLP which passes it through to the codec and on to the IP set. The
difference between the received signal and the echo artifacts is now so small it is not discernible
to the caller.
Here are the requirements for successful training of the adaptive filter.
The adaptive filter can only successfully train in "single talk mode." In other words, where the
calling party from the BCM is speaking (transmit) and when there is no speech from the
called party (receive). This would be typical in a normal conversation. If the receive side is
transmitting a constant signal (such as a radio broadcast or music on hold as was used in
some of the lab tests) the adaptive filter cannot identify the echo component of the transmit
side and therefore cannot train properly.
Training is done on a per call basis.
It can take up to a maximum of 10 seconds to train the filter. If double talk occurs during the
training period the training period will be prolonged. The echo canceller algorithm constantly