Human brains aren't performing in a way which is directly comparable to a CPU cycle. My focus is in psychology and molecular biology, so my understanding of computers may be imperfect but my understanding of brains is strong. A CPU takes a large number of instructions, organizes them in the bus, and operates them singly (or based on how many cores the CPU has, working in tandem. A brain has each neuron as a simple CPU, but there are several different types of neurons (four, by one level of classification) with different types of firing rates and different access to metabolic resources (comparable to power supply for a CPU) even for the same type of neuron at different places in the brain. The AI would, like a human, be doing a fractal buttload of processes all the time in parallel, and those processes would invest in influencing system outputs, and those processes would compete for future system resources by the success of their influences, and as networks of processes came up and went away, so the most successful processes would be regulated internally and you'd have the adaptable, universality of a true AI. This is how the brain works.

I posit that it would, like a human, have an auxiliary center for speech processing. Consider the problems:

It would, like a human, feed in that speech processing to its other computations at whatever rate it was willing to do so. For many people, when they hear someone speak and then stop speaking for a moment to think, the mind starts trying to predict what the rest of the words are going to be to have a response pre-processed and ready. Problems with this approach are 1) a predictive process make a prediction based on incoming data what the speech is going to be, spawning a bunch of other processes investigating the possible predictions, and thus investing a lot of resources into one predicted problem and solution, and then having predicted the speech-question wrongly, so that the entire effort was a waste, 2) a predictive process investing a lot of resources into one prediction and then corrupting the parallel process which is overseeing the various predictive processes to cause the overseer process to not cause it to cancel its investment.

Now alternately you could solve this by having the machine dedicate its attention to the message with small dedicated sets of processes which process speech input. These could run either by: processing the individual words and holding them in memory, or as a fractally smaller level of the networks above, so that they were just running several parallel predictive processes to predict the words, and feeding those processes out to the main language processes to the degree that they seemed likely, etc.

The point is that a true AI would not be impatient, because it wouldn't cost an inordinate amount for it to be patient, because it would be dedicating the appropriate amount of resources at each level and wouldn't be sitting around tapping its foot about waiting for this human to stop talking.