Audibility of digital reconstruction filters

[Miska]

Now here's 54 µs long pulses. It is only 19 samples long at 352.8 kHz, 17 samples would have been 48 µs. This can still barely represent the differences in two waveforms. If you convert this down to 44.1 kHz sampling rate it becomes severely distorted and cannot be represented correctly anymore.

 

http://www2.signalyst.com/tmp/gg4-16.flac

 

Still the difference is clear.

 

Here are the spectrums.

 

Looks awfully lot like a filter, doesn't it?

 

Here you can compare to the 44.1k conversion:

http://www2.signalyst.com/tmp/gg4-16-441.flac

[Fokus]

Looks awfully lot like a filter, doesn't it?

 

Yes, these are filters. And your point is?

[Miska]

Yes, these are filters. And your point is?

 

They sound different! I can post couple of resampling filter impulse responses combined to a file later today and we can all listen to the filter snaps and see if we can hear differences.

 

Did you ever try to listen how your filters sound alone?

 

Optimal filter for 44.1 kHz sampling rate would be one sample long. Too bad it doesn't do anything. So rather bump up the sampling rate high enough and shorten the filter until the entire reasonable filter can fit into equivalent of one sample of 40k sampling rate. Or the other way, bump up sampling rate until you find that first order filter is enough.

 

By the way, did you ever check how filters and noise floor converge in DSD? Maybe someone gave it a thought?

[Fokus]

Of course they sound different. One of them is louder in the audible band, and the other is flatter there.

[13mh13]

Not sure I totally like the newer nomenclature of D/A digital filter ... e.g., RECONSTRUCTION instead of older terms like "oversampling" or "interpolation" digital filter. True, by theory, a lot of "reconstruction" gets done in the DF, but the final waveform reconstruction is done in the DAC. Also, in some designs, a multi-pole ANALOG filter -- LPF, or low-pass-filter -- can follow the DAC output

 

IAC ...

 

I wish more user-level DF control options were avail. for computer audio cards. E.g., my Asus Xonar card uses a PCM1792 DAC (that has the DF built inside). The datasheet for the 1792 notes many functions can be software controlled (e.g., 4x or 8x oversampling, etc.). Certain stand-alone D/A processors do offer control. Unfortunately, Asus drivers do not allow much parametric control over the DAC via Asus's Xonar Audio Center software.

 

My 2011 $120 Colorfly ck-4 iPhone-sized DAP, with a Cirrus CS4398 DAC, offers goodies like user-selectable fast and slow roll off digital interpolation filters (this directly controls the Cirrus DAC).

 

There are a few Asus Xonar user groups that write custom drivers, but none have tackled the DAC chip parameters noted above

[Miska]

Of course they sound different. One of them is louder in the audible band, and the other is flatter there.

 

Regardless of the level, the other has notably my high frequency content. So ear can clearly hear the pitch difference from just 50 µs long sound. Converted to 44.1k both also sound notably softer than the original.

 

Returning to the original topic I can hear differences between different types of oversampling filters, even if it's same filter but minimum or linear phase, just by listening to the filter "snap". If the two filters are not exactly the same, they sound different.

 

Of course you can also take one of the many DACs with multiple filters and just listen different filter settings with dirac pulse.

 

For example my Pioneer SACD player has "Legato Link" filter that has early HF-rolloff so based on frequency response it should sound smoother, but to me it sounds more hissy with added sibilance in voices (likely due to higher ultrasonic leakage). So it's unbearable to listen and I have to turn it off.

[Miska]

There are a few Asus Xonar user groups that write custom drivers, but none have tackled the DAC chip parameters noted above

 

My ASUS Xonar DX with CS4398 has selectable filter on Linux. So you could check if other ASUS cards would gain this option too by using Linux.

[esldude]

When listening to the files I attached in a previous post, I heard a distinct difference in a sine wave vs a sine wave with a steep transient at the beginning. I thought it would sound different, but was surprised how different they were. I initially did this at 44 khz sampling and thought ,"hey, maybe there is something to steeper transients being audible". So I then proceeded to do the same thing at 176 khz sampling to see if there was even more difference. I found while the two sounds still sounded different the waveform with a steeper transient sounded no more different at 176 than it did at 44. Makes me think our ability to track transient attacks is no better than our ability to hear sine waves. In other words transients faster than our upper frequency hearing limit around 20 khz (15khz for me personally) sound no different than one limited to an equivalent 20 khz limit.

 

I also have trouble imagining such a waveform would ever result in real music. Such a waveform would occur if musical instruments burst forth energy without starting the wave at zero level. But I fail to picture how that would occur naturally. I also mixed 20 khz waves with the 4410 hz tone. Like one might get with something like cymbals being struck. Lots of high energy overtones that I think people imagine would give a very steep transient sound. When done this way with both tones starting at zero together as would happen with instruments I heard no difference with or without the 20 khz mixed in. Most likely because I can't hear 20 khz. So again this leads me to think transients above my basic upper frequency hearing limit are not so surprisingly not being heard.

 

We can see in the waveform that high energy things like cymbals ring the filters, and that the ringing at higher sample rates gives steeper transients, but it seems likely to me all of that is beyond our ability to perceive. It therefore would not effect the sound we hear. So not sure what the higher sample rates get us beyond needless bandwidth. Maybe 88 or 96 to prevent interactions with tweeter resonances and such. Beyond that it surely appears useless. If other digital filters sound different without effecting basic frequency response I do wonder what it is and even if it is.

[esldude]

They sound different! I can post couple of resampling filter impulse responses combined to a file later today and we can all listen to the filter snaps and see if we can hear differences.

 

Did you ever try to listen how your filters sound alone?

 

Optimal filter for 44.1 kHz sampling rate would be one sample long. Too bad it doesn't do anything. So rather bump up the sampling rate high enough and shorten the filter until the entire reasonable filter can fit into equivalent of one sample of 40k sampling rate. Or the other way, bump up sampling rate until you find that first order filter is enough.

 

By the way, did you ever check how filters and noise floor converge in DSD? Maybe someone gave it a thought?

 

I listened to these Miska and could not hear a difference. If there is one it must be quite subtle and small. I did have to convert your 352 file down to 176 as I don't have a 352 capable DAC. They sound the same to me. The time difference between channels is well within the ability of 44.1 to portray.

[Miska]

When listening to the files I attached in a previous post

 

What kind of equipment you used for listening, does it reproduce to at least to 50 kHz?

 

I also have trouble imagining such a waveform would ever result in real music.

 

I'm not sure what you mean, but maybe something like this:

instruments.zip

 

Note, these are normalized but not compressed, so I have to dial up my volume control about 20 dB higher than normal to get close to real world SPL. (I think the instrument's sound is somewhat above 100 dB SPL, but I didn't measure it yet)

 

(you may need to play this in loop in case your DAC's start-mute eats beginning of the transient)

 

So not sure what the higher sample rates get us beyond needless bandwidth.

 

I rather record lossless, IOW, that ADC, digital domain or DAC doesn't restrict dynamic range or bandwidth of the original signal.

 

From the above test recordings we can see that the instrument disappears into noise shaping noise around 65 kHz. So 96 kHz sampling rate wouldn't be enough for lossless recording. But with these mics 192 kHz is. This still doesn't have clarity of the live sound, so better microphones and headphones would be needed.

 

If I'd have for example the Sanken microphone probably 192 kHz sampling rate wouldn't be enough for lossless recording.

[Miska]

The time difference between channels is well within the ability of 44.1 to portray.

 

Time difference is not so much problem for 44.1k. Theoretically RedBook can represent timing differences up to 350 ns for continuous tones, given that DAC would have perfect linearity and SNR to -96 dBFS. Practically it's somewhere closer to 1 µs.

 

Time difference play is there just to provide some stereo imaging demo since I generated stereo files.

[sandyk]

Loudspeakers are bad for any hearing tests. I listen primarily with headphones anyway, while working. I have time maybe once a week or once per two weeks to power up my stereo system and listen through speakers... But I can listen many hours per day with headphones. - Miska

 

Miska also uses wide bandwidth amplifiers that are low noise. You really need wide bandwidth low noise amplifiers , (preferably non switching types at the present state of the art) to hear these subtle but very worthwhile differences that bring us closer to the original sound. IIRC, Miska's PA is also flat to 100kHz, and he also mentioned previously how good a

previous amplifier with 1mHz ( !) bandwidth sounded. Far too many source components are dumbed down by "measurements are everything" type EEs with draconian output filtering in order to be able to boast about the ultimate THD figures achieved, which is invariably at the expense of achieving a more natural sound. The same applies to the use of retro DSP which will almost certainly degrade/remove these subtle improvements.

[esldude]

What kind of equipment you used for listening, does it reproduce to at least to 50 kHz?

 

 

 

I'm not sure what you mean, but maybe something like this:

[ATTACH]11660[/ATTACH]

 

Note, these are normalized but not compressed, so I have to dial up my volume control about 20 dB higher than normal to get close to real world SPL. (I think the instrument's sound is somewhat above 100 dB SPL, but I didn't measure it yet)

 

(you may need to play this in loop in case your DAC's start-mute eats beginning of the transient)

 

 

 

I rather record lossless, IOW, that ADC, digital domain or DAC doesn't restrict dynamic range or bandwidth of the original signal.

 

From the above test recordings we can see that the instrument disappears into noise shaping noise around 65 kHz. So 96 kHz sampling rate wouldn't be enough for lossless recording. But with these mics 192 kHz is. This still doesn't have clarity of the live sound, so better microphones and headphones would be needed.

 

If I'd have for example the Sanken microphone probably 192 kHz sampling rate wouldn't be enough for lossless recording.

 

I used a Focusrite Forte with Beyer DT880 head phones supposedly having response to 35 khz. My speakers have response to 28 khz. Also used a desktop sound card and a couple other headphones.

 

As for waveforms, when I added high frequencies to a lower frequency so the resulting waveform started more steeply it wasn't heard if the upper frequency was above my hearing range. This is quite normal and how the 3 files in the zip file you attached are composed. The fundamental and all the harmonics start off at the same time.

 

Now the ones which are unnatural and I can't picture how they would occur were those I made with an artificial jump at the beginning. It went from zero to the peak of the sine wave in one sample period. Those sounded obviously different. But there was no additional change in sound between them if I made the first step more steeply than in 44.1 by using a higher sample rate. Leading me to believe the higher steepness isn't heard. And in any case I don't know of an instrument that would create such a waveform in the air in real life.

[Miska]

The fundamental and all the harmonics start off at the same time.

 

Yes, that's a minimum-phase response. The instrument tones I posted converted down to 44.1 using traditional brickwall gain some pre-ringing due to linear phase filter and start in only couple of samples:

 

If you compare it to a minimum phase FIR filter (this one is at 352800, for 20 kHz cut-off):

 

Now the ones which are unnatural and I can't picture how they would occur were those I made with an artificial jump at the beginning. It went from zero to the peak of the sine wave in one sample period.

 

It will trigger your DAC's oversampling filter ring, but it will smooth the jump to bandwidth so you will get fast rise with ultrasonics up to fs/2 (here upsampled 8x to 1.4 MHz):

 

So if you hear difference between the different starts, you hear difference between different steep rise waveform fronts, because the rise spectrum energy difference is still all just at the beginning:

[13mh13]

My ASUS Xonar DX with CS4398 has selectable filter on Linux. So you could check if other ASUS cards would gain this option too by using Linux.

I looked at the DX's Manual. If you're referring to selectable sample rate, the Xonar ST/STX has this, too.

[13mh13]

In many ways, you folks are carrying on about stuff that's been covered previously at other joints .... perhaps most infamously by Keith Howard at Stereophile magazine, way back in early 2006 (this was in the print version, too)...

 

Ringing False: Digital Audio's Ubiquitous Filter | Stereophile.com

Keith Howard's Ringing False: Digital Audio's Ubiquitous Filter | Stereophile.com

 

Some graphs from the Stereophile article:

 

[esldude]

Yes, that's a minimum-phase response. The instrument tones I posted converted down to 44.1 using traditional brickwall gain some pre-ringing due to linear phase filter and start in only couple of samples:

[ATTACH=CONFIG]11690[/ATTACH]

 

If you compare it to a minimum phase FIR filter (this one is at 352800, for 20 kHz cut-off):

[ATTACH=CONFIG]11691[/ATTACH]

 

 

 

It will trigger your DAC's oversampling filter ring, but it will smooth the jump to bandwidth so you will get fast rise with ultrasonics up to fs/2 (here upsampled 8x to 1.4 MHz):

[ATTACH=CONFIG]11692[/ATTACH]

 

So if you hear difference between the different starts, you hear difference between different steep rise waveform fronts, because the rise spectrum energy difference is still all just at the beginning:

[ATTACH=CONFIG]11693[/ATTACH]

 

Yes, I do hear a difference. If you had listened to or look at the files I attached. However the difference doesn't increase with steeper waveform fronts. Sounds the same with a steep 44.1 khz jump as with a steeper 176 khz jump. Makes me think once the steepness exceeds the max we hear more is of no consequence. When you listened to the file with multiple tinks did all of the tinks after the first one sound the same to you or did each sound different?

[Miska]

I looked at the DX's Manual. If you're referring to selectable sample rate, the Xonar ST/STX has this, too.

 

No, I'm talking about capability to select digital filters of the DAC chip.

 

DX's manual is irrelevant here because it is in no way related to the Linux driver. Which is not developed by ASUS.

[Miska]

In many ways, you folks are carrying on about stuff that's been covered previously at other joints .... perhaps most infamously by Keith Howard at Stereophile magazine, way back in early 2006 (this was in the print version, too)...

 

2006 is not "way back"... I already had HQPlayer at that point and had been developing filters for 10 years...