The challenge within echo cancellation lies in the ability to rapidly track changes in the echo path while remaining robust to large bursts of disturbances, such as double-talk. These two objectives are contradictory because in order to rapidly train to the changes in the echo path, a system requires an adaptive algorithm with a fast convergence speed, which in turn means it will easily diverge in the presence of double-talk.
In most systems, a double-talk detector is implemented to freeze adaptation and prevent divergence. The decision threshold for a double-talk detector generally requires some tuning to the signal processing environment. This is an undesirable characteristic, thence, most systems are tuned to be overly conservative. In other words, since a double-talk decision miss is more important to prevent than a double-talk false alarm, a system is tuned to insure no missed decisions occur. This can result in a system that can produce false alarms at a high frequency, preventing a system from adapting when it should.
The application of robust adaptive algorithms to a echo cancellation system allows the double-talk detector to be less conservative. As the majority of the double-talk misses occur at the onset and offset of speech segments, robust adaptive algorithms attempt to make the adaptation of the filter coefficients impervious to disturbances until a double-talk decision can be made. A scaling factor that is regulated by the level of the past error signals, controls the rate for the adaptive algorithm. Thus, a system that has not converged for example, will result in a high scaling factor. In a condition of double-talk, it’s onset will be delayed as the past error signals values are used.
The Two-Path Method for Acoustic Echo Cancellation (AEC) paper, discussed this topic further. The complications of double-talk detectors and the purpose of using two sets of filter coefficients as an approach to mitigating the problem of fast convergence and double-talk become clearer. The problem with the two-path method is the requirement of two FIRs operations, which imposes a necessary increase in memory and computational complexity to the system. The application of a robust adaptive algorithms can help alleviate the problem of overly conservative double-talk detectors without the added burden of the two-path method.