Only on the final output file, since temporaries are in floating point format. It does not make much difference, though, because the output file still has 24-bit resolution, so it should not really matter in practice if it is dithered or not.
Only on the final output file, since temporaries are in floating point format. It does not make much difference, though, because the output file still has 24-bit resolution, so it should not really matter in practice if it is dithered or not.
I suppose a simple listening test of a fixed non-jittered delay would prove if there are other non-jittery artifacts?
As far as the jitter distributions, real oscillators would exhibit more of gaussian distributions, but at the end it is probably fair to say that all we really care about are the spectral characteristics. In partice, these sources usually follow perhaps a -20dB/dec to perhaps 10KHz and then flatten out to near-white above that. If we wanted to split hairs then probably a more accurate model of the random clock jitter would be a sum of two gaussians with one being a higher amplitude LPF version and second being a lower amplitude white.
I can create new samples if there is still interest in the listening test.
It would be interesting to plot the lower limits of our resident golden eared listener, but it would also be nice if others (including frequenters of the more subjectivist subforums) could do some of the tests. I've already bombed on the high jitter tests but am happy to try again with other samples
Regarding the distribution of the noise, due to the 4 Hz lowpass filter, it already had a close to Gaussian distribution, but it did not have the white component. Uniform distribution white noise converted to brown noise with an integrator (-6 dB/octave slope over the entire frequency range) has Gaussian distribution.
Regarding the distribution of the noise, due to the 4 Hz lowpass filter, it already had a close to Gaussian distribution, but it did not have the white component. Uniform distribution white noise converted to brown noise with an integrator (-6 dB/octave slope over the entire frequency range) has Gaussian distribution.
Sure. Probably only the resulting spectral density is the more relevant parameter. The only thing is that uniform distribution (and so is the filtered version) is bounded, where as a true gaussian is not. May be it's easy to approximate pseudo-Gaussian by summing a whole bunch of randomly seeded uniform before applying the LPF?
It would become unbounded with a filter cutoff frequency of 0 Hz (basically, an integrator, which does in fact sum an "infinite" number of previous samples). However, in practice, Gaussian distribution noise has an extremely low chance of having a very high peak amplitude. I tried filtering a sum of 16 uniform distribution generators, but it only extended the "tails" of the distribution slightly compared to applying the 4 Hz lowpass filter to simple uniform noise. Anyway, if someone is interested in trying more ABX tests, the next samples will use Gaussian noise, and more importantly have a white noise component added.
Probably. For instance, for a couple of minute sample at your sample rate is going to generate about 11x the rms value. For example, injecting 1ns rms jitter of Gaussian jitter you'd should, on average, get a bit over 11ns of peak-to-peak within that same window of time. Obviously, well over 99% of the jitter within that window of time is likely to be within the 6ns p-p.
UMS writes - "Like all the tests, this ABX requires replication, including multiple testers, before there's any point drawing conclusions about it."
Valid ABX testing does not include any "testers".
From Wikipedia: "all test results must be counted in order for the result to be valid. This include previous failed tests, which might not be made public, while the successful ones are, or repeated tests. All tests performed should be summed, and the p value calculated from the sum, not an individual test"
http://en.wikipedia.org/wiki/ABX_test
10-12 is such a low number of individual tests that given enough time you can beat 95% confidence level by just randomly choosing X or Y.
UMS writes - "Like all the tests, this ABX requires replication, including multiple testers, before there's any point drawing conclusions about it."
Valid ABX testing does not include any "testers".
From Wikipedia: "all test results must be counted in order for the result to be valid. This include previous failed tests, which might not be made public, while the successful ones are, or repeated tests. All tests performed should be summed, and the p value calculated from the sum, not an individual test"
http://en.wikipedia.org/wiki/ABX_test
10-12 is such a low number of individual tests that given enough time you can beat 95% confidence level by just randomly choosing X or Y.
By and large I agree however even if one person reliably and consistently passes the test it is an interesting finding. Subjects whose performance is a long way outside the mean (better/worse) we normally call outliers, sometimes these are excluded but you have to have really strong justification for this or you get accused of cherry picking. Sadly none of this cuts any ice in the subjectivist forums where discussions about jitter continue (55 posts in the last month alone discussing jitter - not in the Science Subforum) and are unencumbered by empirical testing , if jitter is really such a big deal why not come over here and see how much jitter you can actually hear ?
Before establishing new limits to jitter audibility, UMS has to replicate these results in a controlled test environment.
I guess you could give him identical files and he'd post successful ABX logs...
ime, what he says is very true about jitter. people who've never tried anything better than their laptop as a source, are not aware of jitter.
He talks about everything except the thresholds of perception and the size of the error. Can a human perceive picoseconds as "timing"? No. It's way way way too small. Why the heck does he start talking about drummers? That's like equating a drop of water to the ocean.