This master thesis presents a new efficient method of acoustic echo cancellation targeted at speech recognition for robots. The proposed algorithm features a new double-talk detector, an enhanced initialization and a new noise estimation method. The DTD algorithm is based on the normalized cross-correlation method, uses noise power estimation to be more robust in noisy environment and reacts more accurately to double-talk. The new initialization method switches between two different DTD algorithms to prevent problems during filter convergence. The simple, yet robust Geigel DTD is used during adaptive filter convergence, whereas the program switches to the newly developed DTD after convergence. Finally, the new noise estimation algorithm relies on the output auto-correlation to correctly estimate the noise. To improve speech recognition performance, center clipping is applied on the output of the echo canceler, to further remove the residual echo. White noise is also added to the output signal, in order to make the signal power more stable, which helps the speech recognition engine. Evaluation of the proposed algorithm has been done on a large set of sequences and results have shown that the new algorithm can increase the word recognition rate by up to 80%.