So I created two new test sound samples, based on measurements performed in an anechoic room with a B&K type 4100 head and torso simulator and a rotating table, in front of an Electrovoice loudspeaker.
I assumed that, as in original Blumlein work, the virtual microphones employed for amplitude-panning the signals are two figure-of-eight, crossed by 90°, and pointing respectively at +/-45° from the nose direction.
So I built the Blumlein-panned signal by mixing, with the proper gains, the two measured binaural impulse responses measured with the test loudspeakers placed at +/-45° (this is a stereo playback system with loudspeakers at
+/-45°, which should be the correct value if you want that there is no
angular error at least for the signals panned hard-left and hard-right).
The "proper gains" cited above are obtained respectively as cos(Alfa-45) and cos (Alfa+45), being Alfa the intended direction of the virtually-panned sound, in steps of 5°, and ranging from 45° Left to -45° Right.
The test sound created this way was named Blumlein-All.wav.
This is a binaural recording obtained with the dummy head, in front of a pair of loudspeakers at +/-45°, in an anechoic room.
For comparison, I assembled a collection of binaural impulse responses corresponding to a true loudspeaker placed at each angular position in front of the dummy head.
This file was called Binaural-all.wav .
This way it is easy to compare the "true" sound of a single source being rotated in front of the listener to the "virtual" sound obtained by stereo playback over a loudspeaker pair with the Blumelein panning law.
It is quite interesting to listen at these two WAV files by headphones, everyone can perceive the difference between the true source moving, and the virtual source being Blumlein-panned.

Then I switched to measurements, and I measured the interaural phase for any direction of the real source and of the virtual source. I created an Excel spreadsheet, containing all these results, and some graphs showing the comparison between the phase response of the real source being moved around, and of the virtual source being panned over the loudspeaker pair.
The results show clearly that there is a significant deviation between the phase responses. Furthermore, the error increases with frequency, and becomes quite unacceptable just above 500 Hz. At 1000 Hz, Blumelin is gone already completely crazy....

Please notice, these results come from experimental measurements in an anechoic room. In a real room, the things will be even worse. And going slightly off-axis, even worst again.... And with an humen head, which is certainly less simmetric and less phase-matched than a B&K dummy head, you can just imagine...

In conclusion, the widespread assumption that Blumlein stereo reproduces correct interaural phase in the range 100 to 1500 Hz is proven to be wrong.
I have other sets of measurements, with different dummy heads and with human heads too. As soon as I find some more time, I will process also these measurement results, but I do not expect that things can go better....

Angelo Farina, 2 May 2006