Double Talk Detection (DTD)

There is an additional complication involved with echo cancellation I was neglecting in the adaptive filter section. We don't want to adapt the filter when there is near end speech. Any interference in the input will cause the filter to diverge. And algorithms that converge quickly tend to diverge quickly as well. So there needs to be a way of detecting near end speech. This is easy when there is no far end speech. But when both speakers are talking it's more difficult. Most detection schemes calculate a value based off the reference and input signal, and then compare it to a carefully selected threshold. Then a decision is made to either turn adaptation on or off. Even while adaptation is off, the echo canceller will still function so long as the room model does not change. A comparison of some DTD methods are below. A good detection scheme will have low probability of miss and a low probability of false detection. The plot is called a Receiver Operating characteristic Curve (ROC) [BGMSG].

The detection method I'm using is the Cheap Normalized Cross Correlation Method[Ah][BMC]. It requires very little computation and works well compared to other methods. Unfortunately, the cheap version uses the estimate of the room response to calculate the detection statistic. This makes it difficult to choose a threshold because the detection statistic depends on the room response and the room response depends on the detection statistic. The results are slightly better (and the range of valid threshold values is wider) if we make the rate of adaptation proportional to the detection statistic. Below is a simulation of this method combined with the FAP.