Can you hear the difference between 16bit and 24bit audio files?

[Paul R]

The images above 22kHz are always implied, they are a firm property of a train of samples spaced 1/44100 seconds apart. These images grow real once the signal space is expanded, i.e. the spacing reduced, insofar the larger space encompasses the images.

 

Teresa was talking about converting CD to DSD. If you think that such is done without steep anti-imaging filtering at 22kHz then you are in for a massive

disappointment. And now please think before you make an even greater fool of yourself.

 

 

Wow - talk about an arrogant ass. You take the cake, Mr, Fokus - whomever the hell you might really be.

 

Why don't you show us poor fools how *you* would convert PCM to DSD - post your algorithm even? You can even just post it in English.

 

I would post mine but as I am such a fool, I don't want to be shown up by your Lordship there. I personally am only interested in the data conversion myself. I admit I am sure you know a hell of lot more than I do about that. I have only got a few decades of experience in doing those kinds of conversions, so I am still learning myself.

 

-Paul

[JohnSwenson]

Thanks, it is helpful. I think some of the confusion for me stemmed from the fact that I had always seen the terms "oversampling," "upsampling," "interpolation," etc., in conjunction with filtering, so I assumed the former automatically meant the latter. One thing: I could use some help in clarifying why the aliasing in the upsampling-plus-filtering example is now taking place around 44.1kHz rather than 22.05kHz. If you're not filtering, you're not changing the response, so the aliasing around 22.05kHz stays. If you *are* filtering, OK, the aliasing around 22.05 is ameliorated to whatever extent the filtering dictates. But does the aliasing at 44.1kHz when filtering is done to the upsampled bitstream come from the upsampling, the filtering, the interaction of the two, or none of the above?

 

Yes, if you oversample but don't filter, the aliases are still around 22.05. If you DO perform filtering so as to attenuate the aliases from 22.05 to 44.1, you STILL have alias generated by the NEW sample rate (88.2) which will mirror around 44.1. If the DSP does not completely eliminate the 22.05-44.1 aliases you will still have them at whatever attenuation PLUS the ones generated by the oversampling sample rate.

 

Things start looking rather interesting in a spectrum plot.

 

John S.

[Jud]

Yes, if you oversample but don't filter, the aliases are still around 22.05. If you DO perform filtering so as to attenuate the aliases from 22.05 to 44.1, you STILL have alias generated by the NEW sample rate (88.2) which will mirror around 44.1. If the DSP does not completely eliminate the 22.05-44.1 aliases you will still have them at whatever attenuation PLUS the ones generated by the oversampling sample rate.

 

Things start looking rather interesting in a spectrum plot.

 

John S.

 

Thanks again, John. On the way out to my car, I thought of it like this: You have to have both oversampling and filtering to get aliasing around a new fs/2, the oversampling to get the higher sample rate, and the filtering to get a new waveform to sample at the higher rate. Your better explanation also notes that aliases of the old sample rate will hang around to whatever extent the filtering does not get rid of them.

[Teresa]

...Teresa was talking about converting CD to DSD...

 

No! I converted my legally obtained 24-bit WAV files (88.2, 96, 176.4 and 192kHz) to 5.6MHz DSD and I deleted my 24/96kHz WAV files of physical formats I created with Audacity and then sold (SACDs, DVD-Audios and LPs) since I discovered one is not supposed to retain copies of music they have sold.

 

While the DSD conversion sounds better to me than the original 24 bit PCM music files, they still don't have the effortless, natural, musical sound of DSD from well made analog or pure DSD masters. You guys can debate on the reasons why, I don't care myself as I am only interested in the music. This is why going forward I'm only purchasing DSD downloads from analog or pure DSD masters.

 

I abhor the strident, cold, uncomfortable sound of CD and 16/44.1kHz PCM, heck I don't even care for 24/48kHz PCM. So Fokus never associate me with having anything whatsoever to do with CD or 16/44.1kHz PCM OK?

[Jud]

Speaking of DSD - I've never been quite clear how sigma delta modulation relates to the topics of filtering, resampling, conversion, etc. Is this a different kind of filtering? A different process altogether?

[wwaldmanfan]

I abhor the strident, cold, uncomfortable sound of CD and 16/44.1kHz PCM, heck I don't even care for 24/48kHz PCM. So Fokus never associate me with having anything whatsoever to do with CD or 16/44.1kHz PCM OK?

 

The wonderful music which I grew up listening to in the 1960's, first on a $20 Lloyds portable record player, then my first hand-picked component stereo system, which I bought at Lafayette Electronics with $150 of my bar-mitzvah money, changed my life. I've been a music lover and collector ever since--45's, LP's, cassettes, CD's, and now, digital downloads.

 

I understand the pleasure you must get from enjoying the highest quality music reproduction possible, especially when your body starts to fail you--I've had a chronic illness for the last ten years myself--but, it seems to me, that there is a whole universe of great music out there that is simply not available in hi-res, so it saddens me greatly that you might be missing out on it.

[Fokus]

No!

 

OK. It started with your "I think it might be the elimination of the sharp filters PCM uses." Not being totally aware of it being you, with your particular taste, I took it from CD to DSD. My mistake. But this doesn't change a thing: if converting, say, 88.2k to DSD, one needs a filter at 44kHz. The same filter as usual, which is clearly not 'eliminated'.

 

I abhor the ....

 

Have you ever thought of installing a decent parametric equaliser? It would perhaps give you access to more music.

[Fokus]

Wow - talk about an arrogant ass.

 

From you, I take that as a compliment.

[Miska]

Speaking of DSD - I've never been quite clear how sigma delta modulation relates to the topics of filtering, resampling, conversion, etc. Is this a different kind of filtering? A different process altogether?

 

I would say it's a different process. A way to produce PDM/PWM from something else, source can be analog or digital.

 

If you go from low rate digital PCM, you need to do some oversampling/upsampling first, because SDM is not a "Nyquist sampling system", but always oversampled.

 

Likewise, going from SDM to PCM you do decimation to return to a "Nyquist sampling system".

[Fokus]

The Philips SAA7320 from 1989 or so was their first bitstream DAC, giving rise to a technology vein that would later be rebranded 'DSD'. If I remember correctly it converted 44.1k/16 to a 11.289MHz 1-bit stream (i.e. DSD256 of whatever the numberspotters call it these days) in the following cascaded steps:

-x4 oversampling with reconstruction filter (at 22kHz)

-x32 linear interpolation

-x2 data replication

 

At this stage the data stream was at 17 or 18 bit wide, to be subsequently truncated to 1 bit with second order (third?) noise shaping, i.e. the process where the truncated part is fed back to the input through several loops and with spectral shaping so that its information does not get lost, but rather is distributed/smeared over the multitude of high-rate 1-bit 'samples'.

 

 

Incidentally, the first application of that 7320 DAC was in a rare Sony integrated amplifier, which today in some circles has near-mythical status as containing the best-possible way for playing back CDs.

[Miska]

-x4 oversampling with reconstruction filter (at 22kHz)

-x32 linear interpolation

-x2 data replication

 

The "data replication" also known has SAH (Sample-And-Hold). This kind of process will produce ugly looking output spectrum, but is necessary on those chips due to lack of processing resources.

 

Today most DAC chips use this process:

- Cascade of three 2x FIR oversampling filters (*)

- Sample-and-hold 16x

 

Followed by low-order modulator, typically 3rd order.

 

This is like AD1955, TI chips, AKM chips, etc.

 

(CS4398 seems to use linear interpolation upwards from 8x)

 

If you look for example at my blog post you can nicely see that with PCM inputs the chip has proper oversampling digital filters only up to 8x rate (352.8/384) and uses SAH upwards from there. So you see images (aliases) around every multiple of this 8x rate. This is strongly correlated error/distortion. Compare it to the DSD128 upsample where such thing doesn't happen at all and there's only some remaining uncorrelated noise-shaping noise left.

 

(*) All three stages used for 44.1/48 rates, two stages for 88.2/96 and one stage used for 176.4/192. Each stage going higher has filter about half of the taps than previous stage.

 

At this stage the data stream was at 17 or 18 bit wide, to be subsequently truncated to 1 bit with second order (third?) noise shaping

 

Since they used a low order modulator they needed very high rate, but still low order modulator is susceptible to noise modulation and idle tones.

 

When I've been working with my modulators, making fifth order modulator noise modulation and idle tone free was challenging. While seventh order modulator was much easier to make clean.

[beanbag]

I won't comment on your attitude and assumption of authority in this area.

 

 

Well that's good...

 

Wow - talk about an arrogant ass. You take the cake, Mr, Fokus - whomever the hell you might really be.

 

I would post mine but as I am such a fool, I don't want to be shown up by your Lordship there.

 

...oops

 

This is why we can't have civil discourse.

[Jud]

Well that's good...

 

...oops

 

This is why we can't have civil discourse.

 

I think it's time to move on from "meta" comments altogether (comments about comments, other commenters, etc.) and get back into the very informative substantive discussion that's broken out lately.

[Jud]

The "data replication" also known has SAH (Sample-And-Hold). This kind of process will produce ugly looking output spectrum, but is necessary on those chips due to lack of processing resources.

 

Today most DAC chips use this process:

- Cascade of three 2x FIR oversampling filters (*)

- Sample-and-hold 16x

 

Followed by low-order modulator, typically 3rd order.

 

This is like AD1955, TI chips, AKM chips, etc.

 

(CS4398 seems to use linear interpolation upwards from 8x)

 

If you look for example at my blog post you can nicely see that with PCM inputs the chip has proper oversampling digital filters only up to 8x rate (352.8/384) and uses SAH upwards from there. So you see images (aliases) around every multiple of this 8x rate. This is strongly correlated error/distortion. Compare it to the DSD128 upsample where such thing doesn't happen at all and there's only some remaining uncorrelated noise-shaping noise left.

 

(*) All three stages used for 44.1/48 rates, two stages for 88.2/96 and one stage used for 176.4/192. Each stage going higher has filter about half of the taps than previous stage.

 

 

Since they used a low order modulator they needed very high rate, but still low order modulator is susceptible to noise modulation and idle tones.

 

When I've been working with my modulators, making fifth order modulator noise modulation and idle tone free was challenging. While seventh order modulator was much easier to make clean.

 

"Order" and "taps" and length - Googling shows people speaking of taps and length (number of input samples over which the filter operates) as the same quantity, and "order" one less than the number of taps or filter length. Is this the same sense in which you're using these terms?

[Fokus]

This may help re taps and length:

 

FIR Filter Basics | dspGuru.com

 

Order: while a filter indeed has an order, as a specifying parameter it makes more sense for analogue filters (lowish order); digital filters for audio are of such a high order that the parameter itself becomes uninteresting, people directly cutting down to the more relevant performance parameters such as stopband rejection and passband ripple.

 

What Miska means with his 'order' is the number of feedback loops in a delta-sigma modulator, i.e. the number of times the chopped-off portion of the payload signal is recirculated to the input.

[Jud]

This may help re taps and length:

 

FIR Filter Basics | dspGuru.com

 

Order: while a filter indeed has an order, as a specifying parameter it makes more sense for analogue filters (lowish order); digital filters for audio are of such a high order that the parameter itself becomes uninteresting, people directly cutting down to the more relevant performance parameters such as stopband rejection and passband ripple.

 

What Miska means with his 'order' is the number of feedback loops in a delta-sigma modulator, i.e. the number of times the chopped-off portion of the payload signal is recirculated to the input.

 

That makes a lot more sense re use of terms both concerning the delta-sigma modulator and something like the adjustable parameters for iZotope SRC bundled with Audirvana Plus. In the latter case, while filter length can be seven figures (up to more than 2 million IIRC), order, according to Alexey Lukin, is ~4x the Steepness setting, which is no more than three figures (somewhere in the hundreds I think, don't know what the max is).

[Miska]

What Miska means with his 'order' is the number of feedback loops in a delta-sigma modulator, i.e. the number of times the chopped-off portion of the payload signal is recirculated to the input.

 

Well, not exactly. Filter order defines steepness of the cut-off. Each order increases the steepness 6 dB/octave (in over-simplified case). For modulator order, this defines steepness of the noise increase. Higher the order, more the quantization noise is pushed up.

[Miska]

"Order" and "taps" and length - Googling shows people speaking of taps and length (number of input samples over which the filter operates) as the same quantity, and "order" one less than the number of taps or filter length. Is this the same sense in which you're using these terms?

 

For FIR it doesn't make sense to talk about order of the filter, because it is not as clearly defined as for analog or IIR filters. Also for analog and IIR filters steepness still depends on the filter type also, different filters have different profile (for example elliptical vs Butterworth).

 

Number of taps or length for FIR tells something about steepness of filter, but it doesn't tell much more. In DAC chips it practically means that response starts rolling off earlier at 2x rates vs Nyquist and even earlier than tat at 4x rates vs Nyquist.

 

In the latter case, while filter length can be seven figures (up to more than 2 million IIRC), order, according to Alexey Lukin, is ~4x the Steepness setting, which is no more than three figures (somewhere in the hundreds I think, don't know what the max is).

 

To me it looks like it's the maximum value, it can be much less specifically depending on your steepness setting and conversion ratio.

 

 

P.S. Halving the taps every doubling of the sampling rate is done because it keeps the required number of calculations per second constant... (because you almost always have just one MCLK frequency for all 44.1-base rates and one for 48-base rates, so you have fixed number of clock cycles per second divided by number of input samples per second -> N clock cycles per input sample)

[Jud]

 

To me it looks like it's the maximum value, it can be much less specifically depending on your steepness setting and conversion ratio.

 

 

P.S. Halving the taps every doubling of the sampling rate is done because it keeps the required number of calculations per second constant... (because you almost always have just one MCLK frequency for all 44.1-base rates and one for 48-base rates, so you have fixed number of clock cycles per second divided by number of input samples per second -> N clock cycles per input sample)

 

Thanks for the explanation of why the number of taps would be halved in the typical cascade. Regarding the iZotope parameters, would I understand you correctly to say that signal may drop off in response to the filter before the max filter length is reached, so the setting is the maximum length but not necessarily what will be used in a particular case? And last, in terms of disadvantages of more taps I can think of available resources (e.g., clock cycles), delay (in circumstances where that might be a problem) - what else? Would better smoothing of response be among possible advantages of more taps? Anything else?

[Miska]

would I understand you correctly to say that signal may drop off in response to the filter before the max filter length is reached, so the setting is the maximum length but not necessarily what will be used in a particular case?

 

Yes, that has been my interpretation of the offered settings.

 

And last, in terms of disadvantages of more taps I can think of available resources (e.g., clock cycles), delay (in circumstances where that might be a problem) - what else? Would better smoothing of response be among possible advantages of more taps? Anything else?

 

You can also reduce cascades if you can spare more taps... 4x filter with same response as 2x2 cascade needs twice the number of taps.

 

Given same conversion ratio, more taps means more "ringing" in the filter.

 

For example when HQPlayer talks about "short" filters it is just normalized term compared to "non-short" filter, not about absolute number of taps. In practice meaning gentler roll-off.

[asindc]

I am quite certain that Archimago and I don't share my pair of ears.

[Jud]

Yes, that has been my interpretation of the offered settings.

 

 

 

You can also reduce cascades if you can spare more taps... 4x filter with same response as 2x2 cascade needs twice the number of taps.

 

Given same conversion ratio, more taps means more "ringing" in the filter.

 

For example when HQPlayer talks about "short" filters it is just normalized term compared to "non-short" filter, not about absolute number of taps. In practice meaning gentler roll-off.

 

Ah, OK, that's what "short" implies re the function of those filters. Re reducing cascades - Any reason(s) to minimize or eliminate cascades, or are 2x2 and 4x equivalent as far as you are concerned?

[Paul R]

I think it's time to move on from "meta" comments altogether (comments about comments, other commenters, etc.) and get back into the very informative substantive discussion that's broken out lately.

 

No problem from me - I am probably being irritable because I have little tolerance for incompetence masquaradeing as authority and knowledge.

 

More so when I am pushed. Think I will disappear for a few days. Most of this conversation is filled with a lot of good information. Keep up the good work.

 

-Paul

[Miska]

Ah, OK, that's what "short" implies re the function of those filters. Re reducing cascades - Any reason(s) to minimize or eliminate cascades, or are 2x2 and 4x equivalent as far as you are concerned?

 

Like John said, single pass tends to be better than multipass (cascade/multi-stage). Getting rid of cascades allows minimizing rounding errors. More times you massage the same data, more error you accumulate. It also allows improvement on the overall precision.

[Rexp]

No! I converted my legally obtained 24-bit WAV files (88.2, 96, 176.4 and 192kHz) to 5.6MHz DSD and I deleted my 24/96kHz WAV files of physical formats I created with Audacity and then sold (SACDs, DVD-Audios and LPs) since I discovered one is not supposed to retain copies of music they have sold.

 

While the DSD conversion sounds better to me than the original 24 bit PCM music files, they still don't have the effortless, natural, musical sound of DSD from well made analog or pure DSD masters. You guys can debate on the reasons why, I don't care myself as I am only interested in the music. This is why going forward I'm only purchasing DSD downloads from analog or pure DSD masters.

 

I abhor the strident, cold, uncomfortable sound of CD and 16/44.1kHz PCM, heck I don't even care for 24/48kHz PCM. So Fokus never associate me with having anything whatsoever to do with CD or 16/44.1kHz PCM OK?

I had a similar anti 16/44 attitude (tried many hi-end cd players) until I started listening to downloads like the new Led Zep 2 remaster via the Qobuz app on my Ipad. Try it , would be interested in your opinion.