No multiplication here, I thought I was only looking at the difference in the envelope with and without the stuff past 22.05kHz. BTW SciPy (the manual) says it is computed using the FFT and IFFT. The numerical noise is still -250dB
(guesstimate) which is degraded from the usual 1e-15 of 64bit floats.
Not sure this can be useful.
Look, Ovi, when I first mentioned it I said it was just a thought, to that effect.
Never said there was formal publication on the subject, far as I know there isn't. Of course, lots of commercial trade secrets are never published, especially these days when there is little public funded research. And who cares what 5% or whatever number of people like to hear/not-hear in their audio equipment anyway, most people don't care. You are one of the latter. We all know that.
I know, that was just an example, otherwise I still don't understand how a sequence of linear operators can lead to a non linear result.
Going through the frequency domain for what should be mathematically a time domain operation (the Hilbert transform) makes me double suspicious...
Anyways, I don't know, I should probably go through that SciPy library myself, unfortunately my time budget is limited...
Learn how to read, Ovi.
What I said was:
I said the POST refers to an article, which it does.
I also said that John apparently thought there was some correlation with THD and listening tests. He did appear to express that in the POST I was talking about. Why else would he mention THD as a parameter? Why would he suggest to Jan there should be some follow up?
You got confused when I was talking about post and thought I was talking about the article it relates to. And then you throw a big fit about it like its all my fault you can't figure it out.
Thinking out loud (although I'm sure you already considered this):
The problem with your filtering method (zeroing the FFT) is that the zeros are only really valid at exactly the frequency sampling points. This filter will have non-zero magnitude response between them. Otherwise said, it looks like a stopband that's zero, but it's in fact zero only at the points represented by the bin, not the full width of the bin.
If you plot this zeroing filter response vs. a large N tap FIR filter you'll note that the two overlap only at a finite number of points. Because the finite nature of the beast, you can get as close as you want to the ideal response (minus the Gibbs) but never exactly...
As a test, you may want to zero filter your data (as you did), then do an IFFT. Next, zero pad the data so that you get M more samples per bin, to see what happens between the bins of the original FFT. Do a FFT of the zero padded data and compare it to the zero filtered data frequency response. You'll note zeros at the spots where you put zeros, but there will M-1 non-zero frequency samples in between them, and the ones close to the transition region will not be numerically well behaved. These zeroes fluctuating due to numerical noise could be at the root of the fluctuations you got (outside the Gibbs effect).
Sometimes I really feel bad of not having more hands on experience with these computer based DSP tools... Maybe in another life.
I don't see that in the post you linked and I am also seriously uncomfortable with this being discussed without having asked Jack what he actually meant. I read the comment as showing a regulator can improve THD of the amplifier, not that the regulator HAS THD, which of course it must have as its output is less than its input.
I feel you have taken a comment out of context and used it to justify something it cannot. Jack is a good guy, just ask him.
Now personally, given that a modern DAC is a serious bit of mixed signal design* I'd be afraid to try and mess around with the power supplies as the chances of making it worse would be hight. Last time I looked (which admittedly was a looong time ago) 6 layers with seperate analog and digital ground planes was the recommended minimum for 16bit operation. Admittedly this was for precision analog not audio, but suggests that trying out different regulator designs without an optimised PCB for that regulator might seem a fools errand. But might be my naivitey...
* If you are going for >20bit ENOB at least. I also know that those versed in the art of doing digital radio consider audio frequency to be DC and of little challenge 🙂.
Well THD of the error amp is measureable (or simmable). But given how low THD is for a half decent opamp I fail to see the problem.
I think it’s interesting to test your analysis.
That was exactly the reason why I asked Scott in #172 to HP filter the signal the same way as he did with the FFT using it as a LP filter and then also subtract its time domain envelope from the original signal.
The two calculated delta signals should now be each others complement IMO.
If not, that would be an indication that something was not accurate.
Hans
Last edited:
AKM uses a Jung regulator for AK4499 Reference Voltage. They use NJM7805 for another output stage voltage. Just sayin'
Yes, that's why I trimmed the files to exactly 10sec both FFT's are then in 0.1Hz bins exactly, there is no information lost. The information "between" the bins is entirely contained in the bins above 22050. Maybe I'm missing something but the results with respect to the down sampling are almost exactly the same as what I get from the DAW I have. The idea is sampling at 44100 and doing a 44100 FFT gives exact 1Hz bins, the same is true for any sampling rate.
Last edited:
Understood. My turn to think I’m missing something, don’t you have to take a sample of an integer multiple of both sine periods length?
Perhaps you should describe (again?) precisely your experiment, so others could easily try an reproduce? Sines, frequencies, sample length, sample rate, etc...
Last edited:
Microsoft PowerPoint
Great info. Note the listening test method is that used by F.Toole. It is short term A/B.
By David Griesenger
The perception of mid Freq and high freq IM distortion in loudspeakers and its relationship to high-definition Audio
THx-RNMarsh
BTW Research indicates >80db spl makes many IM products from HF audible. 80 seems to be a threshold.
So, listening at 80 or lower SPL you are not going to hear the IM products typically produced.
my last LP system--
Great info. Note the listening test method is that used by F.Toole. It is short term A/B.
By David Griesenger
The perception of mid Freq and high freq IM distortion in loudspeakers and its relationship to high-definition Audio
THx-RNMarsh
BTW Research indicates >80db spl makes many IM products from HF audible. 80 seems to be a threshold.
So, listening at 80 or lower SPL you are not going to hear the IM products typically produced.
my last LP system--
Last edited:
Mr. Marsh maybe you missed these tests are double blind. Do you actually think these folks do sighted A/B tests and give them any merit? I can't believe your resistance to listening with no idea at all what you are listening to.
Last edited:
- Status
- Not open for further replies.