Mind the Mic

Neural acoustic field methods predict how a sound arrives at an arbitrary point in a room, and are ranked by a log-spectral distance computed at microphone positions withheld from training. This paper measures how large that distance becomes when only the receiver moves, twice over. The first measurement uses fifteen microphones on a 1.82 m bar that capture a single hammer strike at one instant, so the sound is fixed and only position varies. Position alone moves the distance by 0.1376 natural-log units, against 0.0362 from our preprocessing, so three quarters of the effect is the microphone. Under the convention this field uses, striking a different spot on the object moves the metric about as much as moving the microphone does. The second measurement moves to DiffRIR, a published benchmark covering four rooms, reproducing both rows of its results table in all of them to within 0.7 percent before measuring the same quantity in its metric and split. At the held-out displacement of each room, the gap between a test microphone and the nearest training one, two measured recordings differ by 7.9 to 10.1 times the improvement DiffRIR reports, and no pair sampled differs by less. Even so, checked at each position separately, DiffRIR beats its baseline almost everywhere in the reverberant rooms, but loses to that baseline at 26.8 percent of positions in the acoustically dead one. None of this makes any published mean statistically unsound; at any single position, the score changes more with the microphone than with the model.