An Echo Canceller with Frequency Dependent NLP Attenuation

In this thesis, three different post-filtering algorithms for acoustic residual echo attenuation have been studied and simulated to see if the algorithms work in case of network echoes. The simulations were carried out using real recorded speech signals. The results are presented by different plots. The post-filter was implemented in the frequency domain due to the lower computation complexity. The main drawback of this is that an additional time delay will occur, when using block wise calculation. Finally, results of the post-filter and those obtained with the Non Linear Processor (NLP) were compared to see if the NLP can be replaced by the post-filter or not. The second algorithm, the overweighted wiener filter yields a very high attenuation of the Near End (NE) signal if the overweight is too great. This means that some of the NE frequencies is represented in the Far End (FE) spectrum. The third algorithm gives the lowest attenuation in case of no NE noise and modulates the NE noise mostly. The first algorithm gives the best overall performance. The simulation results shows that the NE and FE signals must be fully separated in frequency to make the post-filter working at optimum. As the post-filter modulates the NE noise, comfort noise has been injected and tested with simulations using the third algorithm. The estimation and injection of comfort noise works best when the NE noise is of a fairly stationary nature. As the main requirement is to remove the entire residual echo, the NLP can not be replaced with the post filter as it leaves some residual echo behind