ACK filtering has been proposed as a technique to alleviate the congestion at the input of a slow channel located on the reverse path of a TCP connection- . Old ACKs waiting at the input of the slow channel are erased when new ACKs are to be queued. In the literature the case of a one-ACK per connection at a time has been studied. In this paper, we show that this is too aggressive for short transfers where ACKs arrive in bursts due to the slow start phase and where the TCP source needs to receive the maximum number of ACKs to increase faster its window. We study first static filtering where a certain ACK queue length is allowed. We show how this length needs to be chosen in order to improve the performance. We present then some algorithms that adapt ACK filtering as a function of the slow channel utilization rather than the ACK queue length. These algorithms provide a good compromise between reducing the queueing delay and passing a large number of ACKs to guarantee a fast window increase.