*** Inverse filter of the MIT Medialab measurement setup *** In this directory only two .WAV files are stored: - OPTIMUS.WAV - OPTI-INV.WAV The first is the impulse response of the loudspeaker+microphone used at the MIT Medialab laboratory during the HRTF measurements made on the Kemar Dummy Head. (look at http://sound.media.mit.edu/KEMAR.html) The authors tryed to equalize the response of the transducers through the creation of an approximate inverse filter, but they were not able to create an inverse filter capable of giving a correct inversion both of the phase and of the gain of the response. They tried a "stupid" FFT inversion, making the IFFT of the reciprocal of the FFT of the IR, but obviously the result was unacceptable (the original file can be downloaded from their WWW site). So they made a minimum-phase inversion (withe the Needy and Allen approach), completely neglecting the phase response, and they used this approximate inverse filter for the equalisation of the HRTF data set, stored in the "compact" database. Anyway, they posted also the original, unequalized data set (FULL data base), and the OPTIMUS.WAV original impulse response. I made a true inverse of the OPTIMUS.WAV file, through the Mourjopoulos least- squares technique. I deliberately limited the length of the inverse filter to 1024 points, because the first temptative (16384 points) produced an inverse filter which was almost completely zero, except for a small region in the middle, long less than 1000 points. Anyone can now download my inverse filter, and apply it to the original FULL data base, obtaining properly equalized HRTFs, without the phase alteration present in the COMPACT data base. Obviously for applying it to the IRs, You have to convolve them, but being both very short, a reasonable procedure is the following: 1) zero-pad both the inverse filter (1024 pts) and the HRTF (512 pts) to 2048 points. 2) take the FFT of both. 3) multiply the two FFT spectra (beware, thay are vectors of complex numbers!) 4) take the IFFT of the result 5) throw away the last 513 points (which should be zeroes): the proper length of the convolved signal is in fact equal to the sum of the length of the two IRs minus one.