I perform audio noise measurements using a very high high-quality slow noise ound card and 24bit/96kHz sampling rate. I wrote a low-level application to sample and compute the RMS noise using this approach. The effective "noise bandwidth" is ~ 46 kHz (measured) using this approach.

These noise measurements are not weighted.

My question is ASSUMING A UNIFORM FLAT NOISE SPECTRAL DENSITY up to 20 kHz, I calculate (integrating the functional form for the A-Weight transfer function) that compared to a unweighted value (for a 20kHz noise bandwidth), the A-weighted value is almost exactly 2.0 dB lower. If this correct? I read an EDN article stating the difference was 2.5dB.

I ran a simulation that cuts off a flat noise source at 20 kHz with an 8th-order Butterworth filter. The difference in noise between the A-weighted measurement and non-weighted measurement is 2.2 dB.