Can you give some examples - I've never knowingly heard the effect of an ISO - I would quite like to if I have a CD with some on.
-
Welcome to ASR. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!
Does Phase Distortion/Shift Matter in Audio? (no*)
- Thread starter amirm
- Start date
And a bunch of associated inter modulation distortion in your speakers (and other places)
In an older talk at the PNW section we gave some audible examples. Some DAC's do "stuff", which I believe, but can not determine for sure, is due to overflow in intermediate quantities in the DSP before the delta-sigma conversion. I can't see inside the chip, but the CRUNCH is quite unpleasant, as is the slowly vanishing near-DC offset that appears at the same time.
Greg, look at this site. https://min.sjtu.edu.cn/files/wavelet/3-wavelet bases.pdf in particular the part about "regularity conditions". If you're going to 'stack' filters, i.e. for stepwise upsample, etc, which is rather an obviously necessary thing, you need at least ONE zero/set of zeros at the original (upsampling) or final (downsampling) fs/2 to meet the regularity requirements.
TL
R - ONE set of zeros at the start (for upsampling) or end (downsampling) at pi are all you need, as long as you pay attention. This is an issue for any filter chain or filter bank, and a huge problem for things like PQMF and the like wherein a creepy kind of modulation at 1/2,1/4, 1/8 ... fs/2 emerges rather annoyingly.
OP
- Thread Starter
- #1,225
No I haven't. To wit, I give a pass to DACs that have their maximum attenuation at 24 kHz instead of 22.05 kHz at 44.1 kHz sampling.
An impulse signal as I have explain, is wildly illegal. Using it then showing that the response of the system is not the same as the said impulse, will cause alarm in viewership who doesn't understand the nature of this signal. What is the correct output of the system, will seem totally wrong to them. Take a look at this:
Person looks at the 48 KHz response and immediately thinks that is completely broken. Yet as you well know, that is the correct, mathematical response to a system that is bandlimited to 48 kHz sampling. Going to higher and higher sampling rate allows more basis functions to be integrated and hence, approximation to impulse increases. But that is simply because the bandwidth is increased. That concept, in frequency domain, is easily understood that we don't need that extra bandwidth. Presenting it in time domain though, causes rampant confusion making folks think need Megahertz sampling rates (not realizing there is a cost to that elsewhere).
Above graph is from the marketing material of an ADC/DAC company by the way (who is completely in DSD camp): https://www.merging.com/highlights/high-resolution
That is the direction I saw @pma going and so I objected to make sure he and others understand what they are doing. As otherwise, they sell you down the path of needing every component in the system to have megahertz bandwidth.
Sokel
Grand Contributor
- Joined
- Sep 8, 2021
- Messages
- 10,541
- Likes
- 11,401
@Kal Rubinson uses their gear and I will probably be the next to follow as they are really nice and flexible.
Merging is very highly regarded in pro community.
- Joined
- Jun 24, 2024
- Messages
- 2,194
- Likes
- 3,694
If they're using misleading material like the posted to sell their gear, I for one would avoid them on principle.
Curvature
Major Contributor
- Joined
- May 20, 2022
- Messages
- 1,949
- Likes
- 2,947
Can you please reupload?Whew. That's a toughie. If you have an unwrapped phase plot, first, you want to do 1st order linear regression on it, this removes the "pure delay" component.
Once you've done that, analyze it for rapid phase changes inside 1 ERB or so. That's what will bite you.
Many of you may have seen an old paper that said "one million degrees of phase shift is inaudible" (intraaurally). That particular amount of phase shift was almost pure delay. Ooops.
WeTransfer | Send Large Files Fast
The simple, quick and secure way to send your files around the world without an account. Share your files, photos, and videos today for free.we.tl
listen to the two files. The amplitude spectra of the two are absolutely identical.
Now listen to the two. How they were made is in the .m file.
gberchin
Active Member
I see that in the context of wavelets, subband filter banks, etc. But for the case of a single antialiasing or reconstruction filter, does it still apply?
BTW, here is a more complete explanation of regularity in this context: https://perso.telecom-paristech.fr/rioul/publis/199304blurioul.pdf
Sokel
Grand Contributor
- Joined
- Sep 8, 2021
- Messages
- 10,541
- Likes
- 11,401
As Amir said, its only about the visuals.
Merging's potential buyers are not expected to be consumer market, they are solely a pro company (even their protocol is pro, don't expect the usual connections, they operate in Dante environment)
So is more important to know what we're seeing and the EXACT test parameters, as depending of who's doing it they can vary greatly.
If there is a series of filters, half-band, the "first" (upsampling) or "last" (downsampling) needs to ensure regularity, is what it comes down to.
Keith_W
Master Contributor
Ravenna. Although lately, Merging has updated its drivers so that Merging products are now Dante compatible.
OP
- Thread Starter
- #1,233
For a while, they made a push into consumer market resulting in some amount of sales to high-end consumers. That has faded though from what I can tell.
Alright, then I apologize for misrepresenting what you meant to say.
If an antialiasing filter has its cutoff (half magnitude) at Fs/2, then it is theoretically possible to get an "illegal" impulse from it.
Any CD that hits full scale (which is a lot of them) probably has at least one intersample over. I've not tried to see if I can hear the effect, so I can't really say anything about audibility except that an intersample over should be more objectionable than similar "ordinary" clipping (analog or properly bandlimited digital) due to the aliasing that occurs. The way my main system is set up, intersample overs aren't really a problem anyway as I already have extra headroom for the digital crossover and room correction filters I use.
There are a number of examples posted in this thread here. I'll throw in one more that I just looked at: Bloom by Larkin Poe. Oversampling 16x with a sharp filter that reaches full stopband attenuation (>200dB) at Fs/2 reveals approx. 1400 overs for the whole album with the maximum peak being +0.53dBFS.
Last edited:
I have a 'c' program that analyzes things including level histograms for both samples and samples at 4x oversampling.
Intersample overs are not rare in some compressed stuff, not at all. And when they are put through a perceptual encoder, it's a mess of epic proportions.
gberchin
Active Member
Intersample overs can occur if the sample instants just happen to straddle the true peak in the analog signal. In addition, if any resampling is taking place in the digital domain, they can potentially occur because of overshoot in the resampling filter itself.
Indeed, the most obvious example is the sequence +1 +1 -1 -1 repeated. But it can be worse than that.
boxerfan88
Major Contributor
- Joined
- Oct 23, 2018
- Messages
- 1,335
- Likes
- 1,630
Similar setup ... digital XO (FIR) and room correction (FIR). After putting the content through the FIR filters I will occasionally get overload (>0dBFS) which will be clipped to 0dBFS before being sent to the DAC. I notice when the "overload" area is reached, the music becomes "grungy or gritty".
With -3dB headroom, it got much better, but I still get occasional grungy/gritty sound on some tracks. I wrote a multi-channel monitor & logger to log overloads.
sample log said:
After studying the logs, it led me to -4dB headroom in the FIR filter processing, and ReplayGain setting of "-1dB prevent clipping". So far so good with this setting.
While my post is not solely related to inter-sample-overs, digital filter processing can also cause overload.
/
Is that due to aliasing from overflow, or something else? (I know very little about lossy codecs beyond a vague idea of what they do)
Lossy codecs I've tested pretty commonly produce samples above full scale with commercial recordings when decoding to float. One such example is this track (seemingly a favorite of Gene from Audioholics). The Opus audio, direct from YouTube, produces 146 samples (out of ~21M) above full scale with the highest peak at +1.37dBFS.
Indeed. A simple example is to apply an IIR allpass filter. No change in frequency response, but the modified phase response very often results in clipping with commercial recordings.
Last edited:
A perceptual coder introduces noise, and as a result, must change the peak values. It can be either direction, depending on happenstance, but in general, if a track is hard-flat against 0dB, as a disgustingly high number are, that means about half or so of them get bigger.
This has lots of problem, in that clipping, for instance, creates harmonics, which in digital promptly alias right on back down to baseband. A great example is, say, operating at 48kHz, a 46/3 kHz sine wave, symmetrically clipped digitally. Oh, look, now there's an easily audible 2kHz tone there.
This is not a pleasant effect, and I am obeying Amir's request not to curse violently when I say "not pleasant".
Similar threads
