Its a consequence of how the Discrete Fourier Transform works. You do not get information for all frequencies (e.g. a continuum) but rather at discrete frequencies. 1kHz is probably not one of the ones that are actually represented by data. I am not sure why 991 Hz was chosen, but my guess is that it's probably the closest to 1kHz...
Take a look at the block diagram for how they've built the test equipment - its using an EPROM for the test waveform generation. 991Hz looks to be chosen because its achieved through dividing the 352.8kHz sample rate by an integer, 356. My guess is the EPROM just contained 9bits' worth (512) of addressed data. Using 1kHz would have required a bigger EPROM and longer address counter since dividing 352.8 by 1 doesn't give an integer result.
Usually in measurement of digital audio performance of converters it should be avoided to use frequencies, levels and signal durations where only a few actual codes are used for the sampled signal.
As an example, Finger wrote in his JAES article:
"Notice in particular that for the CD sample frequency of 44.1 kHz the entries for 1 kHz and 997Hz are either the same of are very close for amplitudes at and below - 50dB. In addition the number of codes used is either identical or nearly so to the number of possible codes for these test frequencies and levels.
At durations of more than 10 ms and levels higher than -50 dB the data begin to differ.
A 1 kHz test signal at -30 dB would use only 423 codes, as compared with 2018 codes for 997 Hz during a 100 ms synthesis."
(Finger, Robert A.; Review of Frequencies and Levels for Digital Audio Performance Measurements; JAES, Vol. 34, No. 1/2, 1986 January/February, 36)
Perhaps, it's related to the fact that at an FS of 44.1k there are very close to exactly 44.5 samples in an 991Hz cycle. Which means there are very close to exactly 89 whole samples for every two cycles of 991Hz. So, I suspect that the choice of 991Hz may be an function of which tone frequency, that's near 1kHz, can be best defined by an (nearly) integer number of samples.
Last edited:
I agree with Jakob2 on this issue.
Not taking into account dither, when you use 1000 Hz, the greatest common denominator with 48 kHz is 1 kHz and the greatest common denominator with 44.1 kHz is 100 Hz. That means you get a pattern that repeats 1000 or 100 times per second and that tests at most 48 or 441 of the possible codes (it could even be half of that when the phase relation between sample rate and sine wave is chosen unfavourably).
With dither, the strict periodicity of the pattern is gone and adjacent codes also get exercised, but that's still only at most 144 codes for 48 kHz sample rate.
991 is a prime number and it is not a factor of any of the normal sample rates. Hence, with a 991 Hz signal, the greatest common denominator will be 1 Hz and many more codes can be exercised.
I don't know why they don't use a prime multiple of a small fraction of 1 Hz, like 999.91 Hz, though.
Not taking into account dither, when you use 1000 Hz, the greatest common denominator with 48 kHz is 1 kHz and the greatest common denominator with 44.1 kHz is 100 Hz. That means you get a pattern that repeats 1000 or 100 times per second and that tests at most 48 or 441 of the possible codes (it could even be half of that when the phase relation between sample rate and sine wave is chosen unfavourably).
With dither, the strict periodicity of the pattern is gone and adjacent codes also get exercised, but that's still only at most 144 codes for 48 kHz sample rate.
991 is a prime number and it is not a factor of any of the normal sample rates. Hence, with a 991 Hz signal, the greatest common denominator will be 1 Hz and many more codes can be exercised.
I don't know why they don't use a prime multiple of a small fraction of 1 Hz, like 999.91 Hz, though.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.