It is usually assumed that triangular dither changes quantization distortion into something that's indistinguishable from additive noise, but as far as I know no-one has ever checked that with blind listening tests.
It shouldn't be difficult to create two versions of an audio file, one with triangular dither quantized to, say, 7 bits and one with 11th-order dither quantized to 8 bits (one bit more). The noise levels will then be exactly the same, but the 11th-order version has smaller quantization steps and is statistically closer to true additive noise; the first eleven moments of the error caused by dither and quantization are then independent of the signal, rather than only the first two moments.
Is anyone interested in blindly comparing such files?
It shouldn't be difficult to create two versions of an audio file, one with triangular dither quantized to, say, 7 bits and one with 11th-order dither quantized to 8 bits (one bit more). The noise levels will then be exactly the same, but the 11th-order version has smaller quantization steps and is statistically closer to true additive noise; the first eleven moments of the error caused by dither and quantization are then independent of the signal, rather than only the first two moments.
Is anyone interested in blindly comparing such files?
I got only two responses so far including yours, so maybe moving it is a good idea.
Thanks for the literature references! In the section that you quoted, the sentence "Admittedly, these tests were informal, and there remains a need for formal psychoacoustic tests of this sort involving many participants under carefully controlled conditions" seems a bit vague to me. Is there any information on whether the tests that they did were blind tests and whether there were subjects in the test who expected to be able to hear higher moments?
OK, I've put two versions of a short fragment of I do like to be beside the seaside on WeTransfer:
WeTransfer
They'll be there until 18 November 2017.
If you think you hear a difference or that you might hear a difference, please do an ABX test using abx under Linux or the Foobar ABX plugin under Windows and report the results of all such trials in this thread.
WeTransfer
They'll be there until 18 November 2017.
If you think you hear a difference or that you might hear a difference, please do an ABX test using abx under Linux or the Foobar ABX plugin under Windows and report the results of all such trials in this thread.
Marcel, have you checked (listened to) the files?? Something is going wrong, there is a noise like a hell in both of them. See spectrum.
Attachments
OK, over my doubts, this is my one and only attempt.
Code:
foo_abx 2.0.2 report
foobar2000 v1.3.7
2017-11-11 16:53:51
File A: silverysea7.wav
SHA1: ecd9a04e7b536b6654a75ad7e98556429986c6a3
File B: silverysea8.wav
SHA1: 5a4d1a92ecde5c951aaa0b40ae83c6fc716e9d8f
Used DSPs:
Resampler (PPHS)
Output:
WASAPI (event) : Speaker (USB Sound Blaster HD), 24-bit
Crossfading: NO
16:53:51 : Test started.
16:54:47 : 01/01
16:54:55 : 02/02
16:55:07 : 03/03
16:55:18 : 04/04
16:55:27 : 05/05
16:55:33 : 06/06
16:55:40 : 07/07
16:55:46 : 08/08
16:55:52 : 09/09
16:56:00 : 09/10
16:56:06 : 09/11
16:56:11 : 10/12
16:56:11 : Test finished.
----------
Total: 10/12
Probability that you were guessing: 1.9%
-- signature --
58bbbfc2a8e1991b30fc97ddbc3dd440bb388a15
Last edited:
As a sanity check, I just downloaded the files, subtracted the original signal and determined the RMS level of the noise produced by the dithered course quantization. The result was -39.12 dB with respect to the RMS level of a full-scale sine wave for silvery7 and -39.14 dB with respect to the RMS level of a full-scale sine wave for silvery8. In theory it should be -39.1339 dB for both files.
The peak level of the noise is quite different, though: -32.63 dB for silvery7 and -29.28 dB for silvery8 with respect to the full-scale level (the real full scale, not the RMS value of a full-scale sine wave). The theoretical values are -32.6018 dB and -26.5812 dB, respectively, but that -26.5812 dB will almost never occur because it requires all eleven random processes that make the eleventh-order dither and the round-off error to be at their maximum and in the same direction simultaneously.
I think this shows that the subjective noise level depends on the probability distribution as well as on the RMS value. That's actually why the ITU-R 468 standard prescribes a quasi-peak detector rather than an RMS detector. It also means that we're not going to find out anything about the audibility of the higher-order moments this way, because I don't know how to subjectively level match the noise.
In any case, you didn't complain about distortion, so apparently the dithered quantization error sounded like noise to you for both cases.
Thanks for participating!
The peak level of the noise is quite different, though: -32.63 dB for silvery7 and -29.28 dB for silvery8 with respect to the full-scale level (the real full scale, not the RMS value of a full-scale sine wave). The theoretical values are -32.6018 dB and -26.5812 dB, respectively, but that -26.5812 dB will almost never occur because it requires all eleven random processes that make the eleventh-order dither and the round-off error to be at their maximum and in the same direction simultaneously.
I think this shows that the subjective noise level depends on the probability distribution as well as on the RMS value. That's actually why the ITU-R 468 standard prescribes a quasi-peak detector rather than an RMS detector. It also means that we're not going to find out anything about the audibility of the higher-order moments this way, because I don't know how to subjectively level match the noise.
In any case, you didn't complain about distortion, so apparently the dithered quantization error sounded like noise to you for both cases.
Thanks for participating!
Last edited:
Mooly, did you notice any differences other than the subjective difference in white noise level that PMA reported?
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.