I tend to prefer slow roll offs and even the NOS mode - what does this say about my setup and preference? (Conclusion: back to fast linear)

[MRC01]

That doesn’t really make it a very reliable comparison. Especially if you have the upsampling to over 300 kHz enabled, there should be no audible difference. ...

This matches my experience trying different digital filters. I mentioned earlier that I can hear the difference with certain sounds like square waves or jangling keys, but not so much with music. That was at low sample rates (44.1 kHz). At high sampling rates, I find no audible difference even with square waves.

PS: along these lines, if you want listening practice or training with anti-aliasing playback filters: take a well recorded piece and resample it down to a very low rate like 16 kHz (Nyqust 8 kHz). This will make the different filters easier to differentiate on playback. From this, you can learn what to listen for at higher sampling rates.

 

[pjug]

Absolutely.
"Fast" and "slow" convey almost nothing useful.

Maybe lawyers were involved?

Or maybe this guy came up with it. Steep slope = fast.

 

[AnalogSteph]

Here's a quick "how to deal with different payback and recording sample rates in RMAA for dummies" with 44.1 playback and 192 kHz record as desired here:

1. Set RMAA 6.4.5 to e.g. 24 bit, 192 kHz. Make sure all tests save for impulse/phase response are checked, then use the "Generate WAV" button. If you have levels already set up, you can cancel saving the Calibration signal and just save the Test signal.
2. You now need to be able to play this back at 44.1 kHz. There are several different ways of going about this - you could use an audio player with a resampler in the DSP chain (e.g. Foobar2000 with the SoX resampler plugin), or just use e.g. Audacity to convert to 24/44 or 32/44 first (project sample rate = 44.1k, export).
3. Back in RMAA, choose the Steinberg ASIO device for recording. Click the ASIO button to verify input channel mapping and bring up the ASIO control panel in there to verify that buffer settings are sane (we're not aiming for lowest latency here, 512 samples is just fine). If you have already set levels, uncheck "Adjust playback/recording levels".
4. Now get both your audio player and RMAA ready. Then, initiate a "Recording only" test in RMAA and start playback of your resampled test tone once RMAA prompts you to do so. Note that this may require separate playback and recording devices, which in this case are actually present, but a number of audio devices actually support independent clocks for playback and recording (notably, HDA chips).

Digital filters built into audio DACs use oversampling (synchronous upsampling) with multiple internal stages that are doubling input sample rate at the output for efficiency reasons (insert zero samples every other sample, apply FIR half-band filter, and bam). The Burr-Brown DF1700 (datasheet) is a good example. Due to being synchronous, they will always scale with fs by default. At higher input sample rates chip designers will often eventually make them start shedding the first stage(s), which you will notice in DAC datasheets by filter performance becoming much worse. Some "NOS" filters are little more than a sample & hold. For an overview of what goes into a complete DAC, the datasheet for the old CS4303 is as good as any. The ADC side is pretty neat as you can actually find papers on some.

If you need some brushing up on sampling fundamentals, I wrote a little intro ages ago, but there certainly is no shortage of tutorials out there if you don't like mine with its rustic ASCII art graphics (this guy has some purtier ones).

 

[anphex]

I will try out your written guide(s) at the weekend. This topic is still bothering me.

Today I passively listened a few hours with the slow roll off and shifted back and forth every now and then. Because my mind actually tells me that the attenuation is indeed necessary, that I will only get more artifacts, more distortion and whatever, I really want to follow what everyone recommends.
But the workings of the reconstruction filter is very noticable and for me more unpleasant than nos. Mslow is probably the closest as the intended source as it gets, as the creator heard and mastered in the studio. But with nos it's just like - damn, I really sound like a snake oil addict now - the digital veil comes down. Instruments sound as they're being recorded in my room.

A few probably very unrelatable examples:
The bright sounding instrument thingy (can't remember the name right now) at about 0:20 sounds normal and fine with mslow. With nos it's as you can hear every single tube ring out as if fingers were just brushing against, them despite the measured cutoff at the highs that comes with nos.
https://anphex.one/index.php/s/kb7b7NZzLwKpPwK

And here it's about the same with the upcoming violin at 2:09. Mslow: yup, that's a nice violin alright. Nos: Damn, I could almost see the strings vibrating
https://anphex.one/index.php/s/oJsbZDRnNQ3SNBZ

Also a thing I noticed in nos is that there's less noticable masking. All instruments come through pretty nice and are easier to focus.

I will make a few more measurements to get to the point of this. Especially the impulse response (I know, time domain vs frequency domain and stuff, but still). If those don't get me an objective reason for my preference I will finally admit that I am just seing audio ghosts. And don't worry, I won't escalate this as far as in a certain other DAC discussion thread.

 

[AnalogSteph]

You hear what you hear, but it may not be for the reasons you think...

Even if perceived realism is higher, how do you know that this is actually a more accurate representation of what's on the recording (the old "Circle of Confusion" problem), or whether you aren't actually compensating for a possible flaw in the speaker setup / room / room acoustics combo? You seem green enough to potentially not have room acoustics tackled very much.

A few probably very unrelatable examples:
The bright sounding instrument thingy (can't remember the name right now) at about 0:20 sounds normal and fine with mslow. With nos it's as you can hear every single tube ring out as if fingers were just brushing against, them despite the measured cutoff at the highs that comes with nos.
https://anphex.one/index.php/s/kb7b7NZzLwKpPwK

FWIW, that one sounds like it may have dropped out of some kind of arranger software or fancy keyboard. So, a bunch of samples. (Not very smart to upload it as a WAV btw.)

Come to think of it, these days I tend to be more concerned with "is this a good song?" than ultimate realism. A lot of nonclassical music does not even exist in purely acoustic form anyway, and may not even make any attempt to sound completely lifelike, quite the contrary (take, for example, the indie aesthetic exemplified by the latest Phoebe Bridgers record, and let's not even get started with PC Music artists). Now of course when blasting ol' Ludwig van, I do want my orchestra to sound sort of realistic with peaks being handled appropriately.

(Trying to think of some well-recorded stuff, NPR Tiny Desk Concerts come to mind, or KEXP, or triple j's Like A Versions.)

 

[anphex]

You seem green enough to potentially not have room acoustics tackled very much.

FWIW, that one sounds like it may have dropped out of some kind of arranger software or fancy keyboard. So, a bunch of samples. (Not very smart to upload it as a WAV btw.)

Come to think of it, these days I tend to be more concerned with "is this a good song?" than ultimate realism.

My living room has an RT60 of 0,3. Heavy sofa, thick curtains and four acoustic wall pictures with 5cm thick basotect foam.

And I am not making this very specific example comparision because the songs are "good" but because I've been listening to those hundreds of times already and therefore claim to notice such differences.

 

[solderdude]

I play my music only from Windows 10 with lossless streaming services and smack my playback devices to the highest setting they provide (Gustard 384/32) for example. Thinking "well oversampling can't be bad right?".

This is an interesting part of the puzzle.
The moment you start to upsample everything to >88.2Hz there will be no 'roll-off' in the audible band.
Also not when playing 44.1 because the upsampler already has a 'sharp' (steep) filter in it which is applied.
This means 44.1 using a normal filter will have nothing above 22kHz (24kHz with some filters) but also when upsampled to say 176.4 kHz the audio bandwidth will just be 22kHz with nothing present till 88kHz.
When you put this through a NOS filterless DAC or a filtered DAC the actual output will be the same except for some small amplitude noise far, far above the audible range which no normal transducer will ever be able to reproduce.

 

[voodooless]

Yes, @anphex still didn’t answer the question if he used the upsampling for his tests and measurements.

As I said before: if so, it strengthens the hypothesis that all comes down to the level difference of NOS vs OS mode on the DAC. No way you’ll hear the difference between 384 kHz upsampled NOS vs OS in a level matched blind test.

 

[anphex]

Yes, @anphex still didn’t answer the question if he used the upsampling for his test and measurements.

As I said before: if so, it strengthens the hypothesis that all comes down to the level difference of NOS vs OS mode on the DAC. No way you’ll hear the difference between 384 kHz upsampled NOS vs OS in a level matched blind test.

All my measrements were done with 44,1/16 since my Yamaha Interface bugged out with anything higher within RMAA. With RME I used 16/96 Khz.

 

[voodooless]

All my measrements were done with 44,1/16 since my Yamaha Interface bugged out with anything higher within RMAA. With RME I used 16/96 Khz.

Okay, so these measurements do then not reflect the conditions that you use while listening? In that case they are of little use then.

 

[anphex]

I just wanted to hook up my interface again to play around a little but now it won't power on no matter which cable or port. So much for Steinberg interfaces, that's my second one that broke on me after a few years with little use. Will look for a new one. Any recommendations?

 

[anphex]

Okay, small update. I just wanted to watch some Netflix on my LG TV with the integrated app and the optical output is fixed at 48khz. Pretty early I noticed that the highs were severly damped with NOS that I had to enable a filter. When listening to music via windows the highs aren't probably damped because of the oversampling I am using (currently 176 Khz because some old games bug out when using more), suggesting that @solderdude is very close with his thought.

So when listening with NOS at common sample rates like 44,1-48 the highs are more or less out of the window. The current state of insight looks like this:
DAC sample rates up to 48 Khz - filter urgently needed
96 Khz and above - NOS probably feasible, depends on preference

 

[anphex]

Another thing that came to my mind recently. Could the unfiltered ultra sonic mirror frequencies damage the amp or tweeter? Provided the amp has enough bandwith to amplify those.

 

[voodooless]

Another thing that came to my mind recently. Could the unfiltered ultra sonic mirror frequencies damage the amp or tweeter? Provided the amp has enough bandwith to amplify those.

Unlikely. Tweeter impedance rises with frequency, so while voltage might be high, current is not. Amps should have filters at the input and output to prevent oscillation.

 

[anphex]

Reporting back after a few months of NOS. Now, after comparing it back and forth against MSLOW filter, I noticed a few things:
When facing a very complex signal like layered e-guitars or synths NOS is a mess compared to filtered modes. Seems like the image posted earlier in this thread regarding intermodulation really is audible, and it's bad.

NOS mode has the advantage on songs with lots of quick impulses, everything sounds a bit more responsive and tangible. Everything sounds really close. But this could also be because of added distortion. After comparing a while again, NOS really sounds like a very soft harmonic distortion.

All filtered modes sound further away than NOS.

So yeah, I settled for MSLOW now. Like I should have earlier. But it was an interesting journey.

 

[Veri]

Reporting back after a few months of NOS. Now, after comparing it back and forth against MSLOW filter, I noticed a few things:
When facing a very complex signal like layered e-guitars or synths NOS is a mess compared to filtered modes. Seems like the image posted earlier in this thread regarding intermodulation really is audible, and it's bad.

NOS mode has the advantage on songs with lots of quick impulses, everything sounds a bit more responsive and tangible. Everything sounds really close. But this could also be because of added distortion. After comparing a while again, NOS really sounds like a very soft harmonic distortion.

All filtered modes sound further away than NOS.

So yeah, I settled for MSLOW now. Like I should have earlier. But it was an interesting journey.

My PCM1794A DAC has this as default filter: (borrowed from mansr)

I would never one such faux-filter/no-filter like slow/nos. This, a steep filter with deep attenuation is as perfect as it gets. Why not go for perfection

 

[anphex]

Because if this. Slow seems like a good tradeoff regarding ringing and stuff while still getting a clean reconstructed waveform.

 

[voodooless]

View attachment 164927Because if this. Slow seems like a good tradeoff regarding ringing and stuff while still getting a clean reconstructed waveform.

Probably good to read this again: https://www.audiosciencereview.com/...out-my-setup-and-preference.25290/post-861766

 

[MRC01]

Keep in mind a few facts:

• A single-bit impulse like this highlights some differences between filters but is an "unnatural" test signal that doesn't occur in music.
• Those ripples/ringing before & after the impulse are at or near Nyquist, which is inaudible.
• Slow roll off can mean 2 different things:
  • The filter doesn't fully attenuate by Nyquist, so it leaks HF whose aliases create passband distortion
  • The filter has a wide transition band, which reduces computation/latency and passband ripple
• Minimum vs. linear phase is independent of fast vs. slow roll off, with different effects.
• This gives 3 degrees of freedom, making 8 permutations of basic filter categories:
  • Phase: minimum vs. linear
  • Stop-band: below Nyquist, or above Nyquist
  • Pass-band: below 20 kHz, or higher
  • Transition-band: wide/gradual/slow, or narrow/sharp/fast
    • Note: this isn't a 4th degree of freedom because it's the difference between the pass-band and stop-band
• A proper/correct filter should be:
  • linear phase
  • stop-band at or below Nyquist
  • pass-band at or above 20 kHz
  • transition-band: doesn't matter much, but wider can be better as it is computationally less expensive and may reduce passband ripple

At low sample rates like 44.1 kHz, there's no wiggle room (pun intended). It's hard to implement a "correct" filter with such a narrow transition band, with the computational power available in common DAC chips. That's why we see them stretch the filter stop-band beyond Nyquist to 24.1 kHz. Higher sampling rates give a bit of room for different filter implementations. For example at 96 kHz, do you want a "slow" filter that uses the entire range from 20 khz to 48 kHz as the transition band? Or do you want a "fast" filter that is flat to 40 kHz with a transition band from 40 kHz to 48 kHz? Theoretically it shouldn't matter, both should be correct and transparent. But in reality, with the real-world limitations of actual hardware, it might matter.

 

[MRC01]

Probably good to read this again: https://www.audiosciencereview.com/...out-my-setup-and-preference.25290/post-861766

This example is contrived and unrealistic, because the source signal is heavily compressed and distorted. When the source is that messed up to begin with, how much can it matter what kind of filter you use? Decorating a turd.