Blind acoustic echo cancellation without double-talk detection

The near-end and far-end signals in an echo cancellation problem are modeled as independent Gaussian random processes obeying a structured form of non-stationarity in which the statistics are constant over short intervals but are allowed to change from one interval to the next. With these assumptions the echo cancellation problem is cast as a semi-blind source separation problem and principles of maximum likelihood are applied to estimate the echo cancellation filter. The form of this estimate is analyzed as the width of the intervals decreases to a single sample. This leads to an adaptive filtering algorithm reminiscent of recursive least squares (RLS) but enjoying superior convergence properties due to the presence of a normalizing factor that regulates the flow of information into the statistical quantities used by RLS. The normalization acts intrinsically as a “soft” double-talk detection mechanism. Simulation results demonstrate the behavior of the algorithm during multiple conditions including double talk. It achieves 10-15 dB lower misadjustment than conventional RLS adaptive filtering.