Probability of detection of residual echo based on magnitude-squared coherence estimate

Residual echo is a distorted, partially-canceled, and transient echo of the near-end speech signal returned from the remote network. Velardo and Wynn have shown that residual echo can be detected and reduced by using a nonlinear processing technique based on frequency coherence. In this paper a mathematical approach to evaluate this nonlinear processor and the algorithm for computing the probability of detection using the magnitude-squared coherence estimate are presented.