Skip to main content ↑ ↓

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: From 44.1 to 96, which is the best resampler? (Read 67286 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

From 44.1 to 96, which is the best resampler?

Reply #50 –
Quote from: bandpass on
So this is presumably an argument for not polluting this band with aliasing artefacts?
Yes. Or do you actually think that a company like Pioneer who used a concept called "Legato Link" from the early to the late 90s can be accused of bad engineering? "Legato Link" did nothing else as completely keeping aliasing artifacts.
Quote from: bandpass on
Why stop at 30k?
Other configurations didn´t sound like the original.
Quote from: bandpass on
24: because it's need for headroom (and bottom-room) and to mitigate error accumulation.
96: no good reason other than "just in case" and it's a standard meaning that you can get lots of plugins that work at this freq.
Plugins? Not everyone in the business uses plugins. 1. Most mastering engineers are using hardware, not software. 2. Say that to someone like Michael Bishop from TELARC who always uses the highest resolution possible. The same with Bob Katz (www.digidoo.com).
Quote from: bandpass on
Why not just use the original source?
Because one don´t always has the original source. 
Quote from: bandpass on
Okay, but let's see if we can establish that this is actually useful.
Are you suggesting that the ability to choose something is unimportant?

Look, I don´t want to tell everyone that everyone should do this. I just gave an example of how variable this resampler from iZotope is. This is not the holy grail, it is just a possibilty of doing resampling in a perfect way. As I said before this is also something that this resampler can do. Look here: SRC comparisons (only downsampling).
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #51 –
Quote from: Cavaille on
I just wanted to have the upsampling method which sonically came closest to the original source. In the end I was able to "recreate" about 60-70 % of the original sound. I put "recreation" in exclamation marks because I know that nothing can be recreated. One has to be aware of this.

Statements like this are required to be backed up by ABX testing.

From 44.1 to 96, which is the best resampler?

Reply #52 –
Quote from: Cavaille on
Quote

Apparently this resmapler is a configurable generator of nonlinear distortion.  Not exactly high fidelity

Aliasing is not a "nonlinear distortion"

Apparently you are unaware of the relevant definition of nonlinear distortion.

Quick review - nonlinear distortion is distortion which adds responses at new frequencies that are not present in the input. A 44.1 KHz sampled wave by definition does not contain signals above 22,050 Hz. Your graphics clearly show output signals at frequencies > 22.050 Hz. That is all the evidence that is needed to maike an absolute and incontrivertable statement that your graphics show nonlinear distortion.

Quote
- in fact it´s completely able to be controlled.


Controllable nonlinear distortion is still nonlinear distortion.

Quote
It is nothing more than the mirrored frequency band from 20-20.000 Hz - in my example I mirrored only frequencies from around 15.000 to 22.050 Hz into the band starting with 22.051 to around 30.000 Hz.


You've just said that there is nonlinear distortion in accordance with the generally agreed-upon definition of nonlinear distortion.

http://www.amplifier.cd/Tutorial/Distortio...distortions.htm

http://www.embedded.com/columns/technicali...equestid=515988

http://en.wikipedia.org/wiki/Nonlinear_distortion

http://www.pcavtech.com/soundcards/techtalk/nlinear/

From 44.1 to 96, which is the best resampler?

Reply #53 –
Quote from: pdq on
Quote from: Cavaille on

I just wanted to have the upsampling method which sonically came closest to the original source. In the end I was able to "recreate" about 60-70 % of the original sound. I put "recreation" in exclamation marks because I know that nothing can be recreated. One has to be aware of this.

Statements like this are required to be backed up by ABX testing.


Agreed.

This is also a painfully easy ABX test to perform.

Upsample it twice, once without the > 22.050 Hz content, and once with.

The files should be posted in a public place so others can do their own tests.

Failure to do these simple things is IMO the same as admitting that former assertions were incorrect.

From 44.1 to 96, which is the best resampler?

Reply #54 –
Quote from: pdq on
Statements like this are required to be backed up by ABX testing.

I fully agree. Where can I upload snippets of Audio material? Which format should they have? How big are they allowed to be? They will be 24/96 - quite large.
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #55 –
Quote from: Cavaille on
Are you suggesting that the ability to choose something is unimportant?
...
it is just a possibilty of doing resampling in a perfect way.

I'm sure that for some use-cases, upsampling with aliasing as you suggest might be more perfect than without aliasing; it's just not that clear yet what those use-cases are.

  -bandpass

From 44.1 to 96, which is the best resampler?

Reply #56 –
You've lost me now. How can adding aliasing artifacts make any encoding more "perfect"? Surely adding anything that wasn't in the original source material is less perfect? 

Cheers, Slipstreem. 

From 44.1 to 96, which is the best resampler?

Reply #57 –
Quote from: Cavaille on
I just wanted to make clear that the resampler from iZotope can be used in every way one person likes it to be. You can use it the normal way (without any aliasing at all) or with a lot of aliasing. I love to have the option to choose what I want.
Flexibility is great, but there's a risk too. Although it apparently sounds good to you, I'm still not convinced that your method of filtering is the best way.
It's still not clear to me how your test was set up. You mentioned a 24/96 original, so I assume you first did a SRC to produce a 44.1 version and then tried to upsample in a way to come audibly as close to the original as possible. Is that it ? If so, what SRC was used ?
Quote from: Cavaille on
Say that to someone like Michael Bishop from TELARC who always uses the highest resolution possible. The same with Bob Katz (www.digidoo.com).
Don't forget that this is HA and not Gearslutz. It can be interesting to hear opinions from famous and highly respected engineers. It's just that on HA double blind listening test results are considered more valuable to scientific discussions. Double blind also means that you don't look at the Grammies on the wall

ps: on the first page of this thread I've been corrected by SebastianG for mixing up aliasing and imaging. Since we are clearly discussing imaging here may I suggest to use that term to avoid confusion ?

From 44.1 to 96, which is the best resampler?

Reply #58 –
Quote from: Slipstreem on
You've lost me now. How can adding aliasing artifacts make any encoding more "perfect"? Surely adding anything that wasn't in the original source material is less perfect? 

Cheers, Slipstreem. 

Of course, resampling is never perfect -- it should be`best compromise'.

The premise seems to be that when upsampling from 44.1 to 96k, imaging (  ) is preferable to ringing (though both are inaudible @ > 22k).  However, it's not clear why the upsampling is taking place; e.g. since almost all music is listened to at 44.1 or 48k, will we be downsampling again afterwards?

  -bandpass

From 44.1 to 96, which is the best resampler?

Reply #59 –
Quote from: Kees de Visser on
It's still not clear to me how your test was set up. You mentioned a 24/96 original, so I assume you first did a SRC to produce a 44.1 version and then tried to upsample in a way to come audibly as close to the original as possible. Is that it ? If so, what SRC was used ? Don't forget that this is HA and not Gearslutz. It can be interesting to hear opinions from famous and highly respected engineers. It's just that on HA double blind listening test results are considered more valuable to scientific discussions. Double blind also means that you don't look at the Grammies on the wall

ps: on the first page of this thread I've been corrected by SebastianG for mixing up aliasing and imaging. Since we are clearly discussing imaging here may I suggest to use that term to avoid confusion ?

  You´re right of course. Nice sentence with the Grammies, I laughed very heartily - kudos. And you´re very right, I used a track bought over at www.hdtracks.com from Reference Recordings that natively was in 24/96 (for those interested, Rachmaninoff´s Symphonic Dances) and downsampled it with iZotope RX - this downsampled version was upsampled again. I did the same with a downsample done with SSRC in foobar2000.

I didn´t read about imaging - but I will. It appears to be more fitting, thanks for the advice. BTW, I did some testsamples ready to upload. I extracted a snippet from mentioned Reference Recordings, downsampled it and made two upsamples. Together with the 24/96 they would make up around 30 MB. Could I upload them here at the upload section? Every sample is under 30 seconds in length.
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #60 –
Quote from: bandpass on
Of course, resampling is never perfect -- it should be`best compromise'.

The premise seems to be that when upsampling from 44.1 to 96k, imaging (  ) is preferable to ringing (though both are inaudible @ > 22k).  However, it's not clear why the upsampling is taking place; e.g. since almost all music is listened to at 44.1 or 48k, will we be downsampling again afterwards?

  -bandpass

True, Upsampling never is lossless. For exactly that reason I needed a very good resampler. I just looked at the SRC comparisons at Infinitewave for downsampling and I assumed that the upsampling qualities would be the same. IMO, Upsampling can fool the ear so that it believes to hear an original 24/96 source. I know that this is highly... ehm... daring... this "sound" is something that can´t be recreated by something like an EQ or compressor, exciter or anything else. To my knowledge there is no soft- or hardware that can change the ringing of a digital recording or create imaging. And if I do upsampling and do some processing afterwards I most certainly keep that material in its upsampled state.
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #61 –
Quote from: Cavaille on
I extracted a snippet from mentioned Reference Recordings, downsampled it and made two upsamples. Together with the 24/96 they would make up around 30 MB. Could I upload them here at the upload section? Every sample is under 30 seconds in length.

I think this would exceed your forum limit. Of course, FLAC or wavpack would help, but I think it would still be too much.  So maybe shorter samples, use rapidshare or similar, or the webspace your ISP gives you?

  -bandpass

From 44.1 to 96, which is the best resampler?

Reply #62 –
Quote from: bandpass on
I think this would exceed your forum limit. Of course, FLAC or wavpack would help, but I think it would still be too much.  So maybe shorter samples, use rapidshare or similar, or the webspace your ISP gives you?


Where are the procedures and rules for uploads posted?

From 44.1 to 96, which is the best resampler?

Reply #64 –
Quote from: evereux on
Quote from: Arnold B. Krueger on

Where are the procedures and rules for uploads posted?

Read the text under the Uploads forum section:
http://www.hydrogenaudio.org/forums/index.php?act=idx

It's a good start, but it doesn't cover quotas, nor does it cover the procedure to upload.  Yes, when posting to uploads there is an attach file option, but given my & Kees recent experience, a little more magic is required.

  -bandpass

From 44.1 to 96, which is the best resampler?

Reply #65 –
Quote from: bandpass on
I think this would exceed your forum limit. Of course, FLAC or wavpack would help, but I think it would still be too much.  So maybe shorter samples, use rapidshare or similar, or the webspace your ISP gives you?

I feared something like this. I will do some research where to put media files for download. If I will make them shorter it would be more difficult to find differences - the effect is of course subtle (as one can guess) but still observable. I thought about using LossyWAV for compressing? It will only add quantization artifacts from 25.000 to 48.000 Hz... there we have it again: the Hypersonic Effect. 
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #66 –
The files are online. Look here.
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #67 –
Out of interest, here's what it looks like:



The converted files have some extra HF noise (most obvious around the 18 second mark; probably noise-shaped dither), but ignoring that, I'm afraid that conventional wisdom has it that if you can hear any difference between any of these, then it's most likely because of non-linear distortion caused by a fault in the play-back chain.

  -bandpass

From 44.1 to 96, which is the best resampler?

Reply #68 –
So brickwall filter was applied after filter with imaging (look at 18-25 kHz range at ~18sec)?

From 44.1 to 96, which is the best resampler?

Reply #69 –
Quote from: bandpass on
The converted files have some extra HF noise (most obvious around the 18 second mark; probably noise-shaped dither), but ignoring that, I'm afraid that conventional wisdom has it that if you can hear any difference between any of these, then it's most likely because of non-linear distortion caused by a fault in the play-back chain.

It is noise-shaped dither since the resampling process works internally at 64-Bit floating point and gives out 32-bit floating-point. I had to do a slight dithering. Apart from that I also thought at first that this would be non-linear distortion. However, we can rule this out. I did a try and removed on my upsampled material every bit of music from 20-22.050 with Algorithmix EQ - only the part from around 22.050 to around 30.000 was left. If you slow this down in Sound Forge with a DSP designed for pitching you actually can hear music again by making that high-frequency slower, shifting it to regions where we audibly can hear it again. (very very slow of course, but the music structure is showing).
I believe that the sound differences are mostly caused not by the mirrored, imaged frequencies but more by the changed ringing. With the steep brickwall filtering you´ll have strong pre-and post-ringing (according to theory), with the softer filtering this ringing would be weaker. Sadly, I don´t know how to prove this. This is only my theory - I would be glad if someone would be able to offer different explanations.

edit 04-01-09: Or do you mean analogue non-linear distortion? The playing of these files is with foobar without any DSP - my sound interface doesn´t have a DSP (E-MU 0202 USB) and uses ASIO (which of course I activated in foobar). For monitoring and listening I always use a Sennheiser HD-600 which is powered by a Corda Arietta (the difference is the same with the headphone wired to the amp or directly to the interface).


Quote from: lvqcl on
So brickwall filter was applied after filter with imaging (look at 18-25 kHz range at ~18sec)?
For the original downsampling and the upsampled material without imaging, yes. So, the downsampled material contains at least one brickwall filtering.
marlene-d.blogspot.com

From 44.1 to 96, which is the best resampler?

Reply #70 –
Quote from: Arnold B. Krueger on
Quote

With keeping certain aliasing artifacts on purpose you can serve this hypersonic effect. But only if the resampler you use does a precise calculation and can be configured. The one from iZotope appears to be such a precise resampler.

I would call it imprecise when it is adding nonlinear distortion to the input signal.

Every resampler adds some nonlinear distortion to the signal: aliasing or imaging. iZotope allows trading this distortion for passband flatness and amount of ringing. It can perform as good as any other resampler.

From 44.1 to 96, which is the best resampler?

Reply #71 –
Quote from: Alexey Lukin on
iZotope allows trading this distortion for passband flatness and amount of ringing.
It's probably a wise decision to leave the final choice to the end user since there seems very little consensus about what's best. IME users "fiddle around" with the settings until it sounds ok to them.
I find it rather frustrating that there doesn't seem to be any methodology about the best choice in a given situation (although that's not uncommon to audio ). Could it really be that, assuming there are audible differences, the preferences of subjects differ that much ?
Alexey, I've done quite some iZotope SRC testing and haven't found any distortion in the passband. Can you confirm that the problems start in the transition band ?

From 44.1 to 96, which is the best resampler?

Reply #72 –
The distortion of SRC consists of:
1. Passband non-flatness;
2. Non-linear distortion due to aliasing or imaging;
3. Filter ringing.

They may be audible or inaudible, depending on the position of the transition band and other parameters.

Distortions 1 and 2 can be seen on these graphs of the default RX SRC preset:


(conversion from 96 kHz to 44.1 kHz)


(conversion from 44.1 kHz to 96 kHz)

To me, most of these SRC settings sound very similar. However since many users are convinced that certain modes sound better than others, RX offers flexible control. E.g. some users cannot tolerate any aliasing/imaging, other prefer minimum-phase filters.
 

From 44.1 to 96, which is the best resampler?

Reply #73 –
Quote from: Arnold B. Krueger on
Quote from: Cavaille on

Aliasing is not a "nonlinear distortion"

Apparently you are unaware of the relevant definition of nonlinear distortion.

Quick review - nonlinear distortion is distortion which adds responses at new frequencies that are not present in the input. A 44.1 KHz sampled wave by definition does not contain signals above 22,050 Hz. Your graphics clearly show output signals at frequencies > 22.050 Hz. That is all the evidence that is needed to maike an absolute and incontrivertable statement that your graphics show nonlinear distortion.

I don't think you got that right. Being linear or nonlinear is a property of the some "black box" that alters a signal.  This box could be a lowpass filter (linear) or some "guitare distorter" (nonlinear) or whatever.  Linearity means -- in functional notation for the black box 'f' -- that a*f(x)+b*f(y) = f(a*x+b*y) for every scale factors a & b and for every signals x & y.  It is true, that many kinds of nonlinear distortions introduce frequency components that weren't there before.  But this is not an equivalence.  Take zero stuffing for example: Suppose f doubles the sampling rate by interleving the signal with zeros.  This would produce image frequencies but it is obviously a linear operation.  The same is true for the sampling process regardless of whether there's any aliasing or not.  Anther example would be frequency inversion: Flip the sign of every 2nd sample. This will turn the spectrum upside down.  It's still a linear operation.

Cheers!
SG

From 44.1 to 96, which is the best resampler?

Reply #74 –
Quote from: SebastianG on
Quote from: Arnold B. Krueger on

Quote from: Cavaille on

Aliasing is not a "nonlinear distortion"

Apparently you are unaware of the relevant definition of nonlinear distortion.

Quick review - nonlinear distortion is distortion which adds responses at new frequencies that are not present in the input. A 44.1 KHz sampled wave by definition does not contain signals above 22,050 Hz. Your graphics clearly show output signals at frequencies > 22.050 Hz. That is all the evidence that is needed to maike an absolute and incontrivertable statement that your graphics show nonlinear distortion.

I don't think you got that right.


You get to be wrong. ;-)

Quote
Being linear or nonlinear is a property of the some "black box" that alters a signal.  This box could be a lowpass filter (linear) or some "guitare distorter" (nonlinear) or whatever.  Linearity means -- in functional notation for the black box 'f' -- that a*f(x)+b*f(y) = f(a*x+b*y) for every scale factors a & b and for every signals x & y.


So far so good.

You've already made one logical mistake though. You stopped too soon.  Knowing what something is, does not necessarily put limits on what it is not. Aliasing is not linear.

Much of our basic mathematical thinking is based on the idea that audio signals exist only in the amplitude domain, which is what your little equations are about. In fact, audio signals exist in both the amplitude and time (or frequency domains).  Aliasing happens in the frequency (or time) domains.

The definition of nonlinear that I provided is absolutely orthodox and classical. Anybody who disagrees with it, whether its you, me, or anybody else does so that their own peril.

Quote
It is true, that many kinds of nonlinear distortions introduce frequency components that weren't there before.  But this is not an equivalence.


Right, its not an equivalence and this should be very apparent.  I made a general (and true and generally accepted) statement about what linearity is not. Linearity is not something that adds signals at new frequencies. Lineraity could also not be other things, but it is generally agreed that linearity is not adding signals at new frequencies and adding signals at new frequencies is what imaging does.

Just because the f(x) = ax + bx**2 + cx**3 +... expression that describes aliasing is not totally simple, doesn't mean that it doesn't exist.  As soon as b and c and... become non-zero, the linearity rule is broken.

Quote
Take zero stuffing for example: Suppose f doubles the sampling rate by interleving the signal with zeros.  This would produce image frequencies but it is obviously a linear operation.


Wrong again. Zero-stuffing is not linear, but this may not be obvious to you in the amplitude domain. Analyze it in the frequency domain, and the non-linearity becomes obviouis.

Quote
The same is true for the sampling process regardless of whether there's any aliasing or not.


You think that sampling is linear? Time to hit the books again!

Quote
Anther example would be frequency inversion: Flip the sign of every 2nd sample. This will turn the spectrum upside down.  It's still a linear operation.


Again, you really need to revisit what for me was second-semester electronics, and basic signal analysis.

Believe it or not, audio takes place in both the amplitude and time (frequency) domains