In an AEC configuration as shown in the Fig.1, let us consider u(n), the far-end speech signal and x(n), the near-end speech signal. The desired signal is d(n)=v(n)+x(n), where v(n) is the echo signal generated from the echo path impulse response. The purpose of an adaptive filter W is to find an echo estimate, y(n) which is then subtracted from the desired signal, d(n) to obtain e(n).
When I am implementing the APA adaptive algorithm for echo cancellation, I am observing leakage phenomenon which is explained as follows:
y(n) will contain a component proportional to x(n), that will be subtracted from the total desired signal. This phenomenon is in fact a leakage of the x(n) in y(n), through the error signal e(n); the result consists of an undesired attenuation of the near-end signal.
Because of this near-end leakage phenomenon, there is near-end signal suppression in the post-processing output.
I am handling the double-talk using robust algorithms without the need for DTD.
I kindly request to suggest me how to avoid near-end signal leakage in the adaptive filter output y(n)?
To avoid this problem, there are two methods, using DTD, or by modifying your algorithm in its updates by incorporating the statistics of near-end signal to control the output, as done by many authors (such as Variable Step Size).
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms