1) For examining individual subject data, the AFNI slider bar + color bar combination is the easiest way to implement the two-threshold process. In the "Define Overlay" window, set the "Thr" (threshold) selector to the sub-brick corresponding to the overall F-test in your statistical map (often sub-brick #0). Then, the slider bar (which sets the threshold) controls the first step of the thresholding process. Typically, the slider would be set to a high value to produce a stringent first threshold step (e.g. F>10, p < 10-6). Only voxels exceeding this value would be passed to the second, more liberal threshold step.
To implement the second threshold, set the "OLay" selector to the contrast of interest (in your case, words vs. tones). Instead of using the default ('**') color bar (the area under "Inten") use a color bar (e.g. #9) that contains a region in the middle with no color. Then, voxels that have a value for this contrast that is too low will not be displayed i.e. will be filtered out by the second threshold step. You will need to turn off AutoRange to get full control over the value used for this threshold. For instance, colorbar #9 doesn't display values that are +-0.05 of the full range, so set the range to 40 to filter out values <2 (because t<2 corresponds to p>0.05).

2) For examining group data, the same general principles apply (note that it is also possible to implement two-step thresholding at the command line using 3dcalc or other AFNI accessory programs). For the first step threshold, I would recommend using the "All" contrast, because you want to examine all voxels that show any activation before examining voxels that show a difference between words and tones in the second step. The "StimEffect" contrast is not good to use in a two-step analysis because it already measures the effect of stimulus type--e.g. in your experiments, voxels that show a significant difference between words and tones. This might filter out auditory cortex or other areas that respond strongly (but equally) to both words and tones and would make the second filtering step redundant. One final note: it is possible to change the middle part of the color bar so that instead of showing no color, it shows a unique color (such as green). This lets you easily distinguish areas that respond to both stimulus types but don't show a preference (e.g. primary auditory cortex) from regions that show a significant preference for one stimulus type or another.