Controlling the capacity of a distributed speech recognition system