Yes, but note that in the case Michael is reporting, all filters have double-precision coeffs and data storage. It is only when passing samples between unit generators that the difference lies (either single or double precision is used). Still, I believe that there can be audible differences.
Victor Lazzarini Dean of Arts, Celtic Studies, and Philosophy Maynooth University Ireland > On 6 Feb 2015, at 18:43, Ethan Duni <ethan.d...@gmail.com> wrote: > > Thanks for the reference Vicki > >> What they are hearing is not noise or peaks sitting at the 24th >> bit but rather the distortion that goes with truncation at 24b, and >> it is said to have a characteristic coloration effect on sound. I'm >> aware of an effort to show this with AB/X tests, hopefully it will be > published. > > I'm skeptical, but definitely hope that such a test gets undertaken and > published. Would be interesting to have some real data either way. > >> The problem with failing to dither at 24b is that many such truncation >> steps would be done routinely in mastering, and thus the truncation >> distortion products continue to build up. > > Hopefully everyone agrees that the questions of what is appropriate for > intermediate processing and what is appropriate for final distribution are > quite different, and that substantially higher resolutions (and probably > including dither) are indicated for intermediate processing. As Michael > Goggins says: > >> In my own work, I have verified with a double-blind ABX comparator at >> a high degree of statistical significance that I can hear the >> differences in certain selected portions of the same Csound piece >> rendered with 32 bit floating point samples versus 64 bit floating >> point samples. These are sample words used in internal calculations, >> not for output soundfiles. What I heard was differences in the sound >> of the same filter algorithm. These differences were not at all hard >> to hear, but they occurred in only one or two places in the piece. > > Indeed, it is not particularly difficult to cook up filter > designs/algorithms that will break any given finite internal resolution. At > some point those filter designs become pathological, but there are plenty > of reasonable cases where 32 bit float internal precision is insufficient. > Note that a 32-bit float only has 24 bits of mantissa, which is 8 bits less > than is typically used in embedded fixed-point implementations (for > sensitive components like filter guts, I mean). So even very standard stuff > that has been around for decades in the fixed-point world will break if > implemented naively in 32 bit float. > > E > -- > dupswapdrop -- the music-dsp mailing list and website: > subscription info, FAQ, source code archive, list archive, book reviews, dsp > links > http://music.columbia.edu/cmc/music-dsp > http://music.columbia.edu/mailman/listinfo/music-dsp -- dupswapdrop -- the music-dsp mailing list and website: subscription info, FAQ, source code archive, list archive, book reviews, dsp links http://music.columbia.edu/cmc/music-dsp http://music.columbia.edu/mailman/listinfo/music-dsp
