I am working on acoustic echo cancellation which has a double talk detector. During the single talk, the ERLE (Echo Return Loss Enhancement) exceeds 100dB and during double talk around 20dB to 40 dB. I have read few papers and saw that for a good echo canceler the ERLE should be between 40-50dB. Am I going wrong anywhere? Is is ok to get such high ERLE?
It is annoying that I replied after 4 years, but this is helpful to others
Yes, ERLE has to be minimum of 40 dB for good echo cancellation. During the presence of only near end speech, you should not calculate ERLE as it will be near to 0dB, so ERLE has to be used only in presence of Far_end signal.
In your case I think you might be using low average filter length for calculation of ERLE. It mainly depends how you are calculating ERLE, ideally if you have ERLE 100 dB after convergence of your adaptive filter, then it is good enough.
The ERLE depends on many factors, such as the length of the echo in samples, the length of adaptive filter in samples, the strategy for residual echo estimation and nonlinear processing.
Usually, the length of the echo is about 10~100ms, more than the length of the adaptive filter. For example, 16kHz sampling rate， 10~100ms is about 160 ~ 1600 samples. But the adpative filter is only 160 ~ 640 samples because of the computation overhead limitation. Considering the length limitation, the adaptive filter can only contribute 20~30dB ERLE. In order to obtain more ERLE, good residual echo estimation and nonlinear processing are also required.
However, 100dB ERLE seems strange. For a 16-bit system, the dynamic range is only 96dB. Hope I can get more insights from the question.