I need to learn about up-conversion and digital filters

[Jud]

Is it something that could be approach to be truly fixed using digital filtering in the DAC?

 

A linear phase filter doesn't have group delay. It does, however, have some amount of pre-ringing. Overall ringing of the filter can be minimized by making it less steep (less of a discontinuity, see the cite to the Gibbs phenomenon in my previous note). However, this means increased aliasing. Thus the tradeoffs inherent in filter design.

 

These tradeoffs can be made less problematic as the Nyquist frequency moves further out of the audible range. This is why nearly all DACs use 8x oversampling internally, and the reason for sample rate conversion software.

[Jud]

My guess is that a good starting point would be to set Audirvana to upsample everything to 176 or 192 and turn off both upsampling and the digital filter in the DAC. iZotope performs these functions with 64-bit arithmetic and up to 2 million filter coefficients, far surpassing the precision of the digital filter in the DAC chip.

 

Regarding your iZotope settings, my ears are tin rather than golden, so take the following with a grain of salt:

 

You should increase "max filter length" to the maximum possible value, i.e., 2 million. That maximizes the accuracy of the filter. Lower settings are offered only for slower computers.

 

Not so. Filter length affects the number of "taps," which can be loosely conceived of as the number of times the filter "touches," or acts on, the signal. The higher this setting is, the more/longer the filter acts on the signal, and thus the steeper the cut, all else being equal. As always with a steeper filter, this creates more ringing.

 

You might try a lower steepness value to reduce ringing.

 

Especially after you've made the filter steep by maximizing filter length!

 

I use 50.

 

I have my own homebrewed filter setting at a steepness of 5, but have heard good results with settings up to 200.

 

You also might try pre-ringing = 0, since many people find pre-ringing more audible than post-ringing.

 

That would be unlistenable for me. First of all, there's no free lunch. When a minimum phase filter is used to reduce pre-ringing (the pre-ringing setting is actually a phase setting - 1.00 is linear phase, 0 is minimum phase; the number is a ratio in which the amount of pre-ringing is the numerator and the amount of post-ringing is the denominator), that energy has to go somewhere, and where it goes is into post-ringing. Thus as you reduce pre-ringing the amount of post-ringing is elevated until the sound becomes obviously reverberant. When tweaking pre-ringing, long before I reached the 0 setting it sounded like I was hearing the music being played in an indoor swimming pool. Waayy too much reverb. My homebrew setting is 1.00. (See my notes above re the Vandersteens, phase correctness and time alignment - your Thiels should be similar.)

 

You might try a lower cutoff frequency to attenuate high frequency distortion in early digital recordings made with insufficiently steep anti-aliasing filters. I believe this is what Meridian calls an apodizing filter. Because my tin ears lack response to 22 KHz, I set the cutoff frequency to 0.8, i.e., 17.6 KHz for 44.1 sample rate. I do not notice any treble rolloff on any recordings with this setting. However, I realize that is heresy on this forum.

 

I've tried settings lower than 1.00 Nyquist, but have always returned to 1.00. I fancy I can hear the diminished highs, though this could easily be power of suggestion (I could hear up to about 16KHz last I checked).

 

Edit: Something that may account for what I hear - from general reading I believe this setting is not a "cutoff" but the "sq rt of 2 over 2" point, i.e., where the signal is .707 of the unfiltered level. Therefore the cut would start at a lower frequency, and this point could be seen as the "knee."

 

Lastly, an even more heretical idea: After you get used to the sound under those conditions, try setting Aurdirvana to upsample to 96 KHz instead of 192. 12 years ago, Dan Lavry asserted that most DAC chips sounded better at 96 than 192 KHz because the chips at that time required more settling time to achieve their lowest distortion. Newer chips may have overcome this problem, but it's worth experimenting with.

 

http://lavryengineering.com/pdfs/lavry-sampling-theory.pdf

 

For a number of reasons I have never liked Lavry's white papers on this subject. IIRC, in at least one of these papers he used measurements at 16MHz to derive his conclusions about chip difficulties at high sample rates. As I've noted previously, the entire series of tradeoffs involved in filter design - aliasing versus ringing versus phase behavior - becomes easier to do, and to do well, the higher above the audible range the Nyquist frequency is. This is the reason 8x oversampling was already standardized by the time the first separate DACs were coming out, and it was done by engineers working for the huge audio companies manufacturing CD players and their vendors, not fringe audiophiles. Moreover, if your DAC accepts input at 352.8/384KHz or at DSD rates and you feed it those rates, you've *bypassed* the DAC chip (either its interpolation filtering or altogether), so you can't be stressing the DAC chip with something it's not being called on to do. This in turn raises the point that these chips are built to upsample lower sample rates to 352.8/384KHz, so it would be quite odd if feeding them at half that sample rate would cause them to misbehave.

 

p.s. Given that you've never heard a difference between DACs, please let us know: (1) whether you hear a difference among these settings, and (2) whether optimizing the Teac this way enables it to surpass your Peachtree.

[Jud]

Does it have some cure using Lanczos factor in delta sigma if used at all in the DAC?

 

This is something new to me (though I'm sure elementary to people who actually know something about digital filtering). Very interesting, thanks Krzysztof!

 

 

P.S. Obviously I don't know the answer to your question, sorry.

[wgscott]

Not so. Filter length affects the number of "taps," which can be loosely conceived of as the number of times the filter "touches," or acts on, the signal. The higher this setting is, the more/longer the filter acts on the signal, and thus the steeper the cut, all else being equal. As always with a steeper filter, this creates more ringing.

 

From Damien's documentation:

 

- Steepness: steepness of the transition band of the lowpass filter. Higher steepness will reject unwanted frequencies but cause more ringing in the time- domain and a higher CPU load

 

- Max filter length: controls the memory (and CPU usage) used by the resampling filter. Default value (500k) should be sufficient for most applications. Can be increased for getting vey high quality output for very high sample rates (e.g. DSD downsampling)

 

The filter length description sounds more like what Bob is saying. Also, if you use the "quality" slider, it has the effect of changing this (and other) parameters. The higher the quality setting, the larger the number for max filter length. The highest quality setting on the slider goes up to 1,500,000 (out of 2,000,000). That also is consistent with what Bob was describing.

 

Is it possible you were conflating this with the first setting (steepness)?

[Jud]

From Damien's documentation:

 

- Steepness: steepness of the transition band of the lowpass filter. Higher steepness will reject unwanted frequencies but cause more ringing in the time- domain and a higher CPU load

 

- Max filter length: controls the memory (and CPU usage) used by the resampling filter. Default value (500k) should be sufficient for most applications. Can be increased for getting vey high quality output for very high sample rates (e.g. DSD downsampling)

 

The filter length description sounds more like what Bob is saying. Also, if you use the "quality" slider, it has the effect of changing this (and other) parameters. The higher the quality setting, the larger the number for max filter length. The highest quality setting on the slider goes up to 1,500,000 (out of 2,000,000). That also is consistent with what Bob was describing.

 

Is it possible you were conflating this with the first setting (steepness)?

 

I've read the Audirvana documentation.

 

The max filter length adjustment varies directly with the number of filter "taps," i.e., they both describe how many samples the filter acts on. Here is what Miska has to say about number of taps (http://www.computeraudiophile.com/f8-general-forum/we-dont-need-no-stinking-hi-rez-26580/index6.html#post486315):

 

Number of taps on the filter define how many samples before and after the event are taken into account. Perfect brickwall filter would utilize infinite number of samples. Even a less than perfect one could utilize more samples than length of a song has. This has also various side effects...

 

This is also directly related to Fourier transform, more points you have, better the frequency resolution but worse the time resolution. And vice versa, less points you use, better the time resolution.

 

[Emphasis added.]

 

With more taps (greater the "length" setting in iZotope), one gets "better" frequency operation (meaning deeper cut) and worse time resolution (meaning more ringing).

[goldsdad]

Moreover, if your DAC accepts input at 352.8/384KHz or at DSD rates and you feed it those rates, you've *bypassed* the DAC chip (either its interpolation filtering or altogether), so you can't be stressing the DAC chip with something it's not being called on to do. This in turn raises the point that these chips are built to upsample lower sample rates to 352.8/384KHz, so it would be quite odd if feeding them at half that sample rate would cause them to misbehave.

 

Perhaps it's not the D/A converter chip itself, but the digital input module of the box of electronics which has been found to have an optimal input rate that's less than it's maximum, in some cases.

[Jud]

Perhaps it's not the D/A converter chip itself, but the digital input module of the box of electronics which has been found to have an optimal input rate that's less than it's maximum, in some cases.

 

I've read the Lavry papers and corresponded with a Lavry representative in this forum, and the Lavry thesis isn't nearly that detailed. Have a look at his stuff and see what sample rates his data points come from. (There are three data points as I recall, one at 16MHz.)

 

I have seen some references to DACs that are happier with input of "4x" rates (176.4 or 192KHz) than input of "8x" rates (352.8 or 384KHz), and that makes some sense to me, since until fairly recently many DACs did not accept inputs over 4x rates, and perhaps the same input circuitry is being used in the new products. But I don't think the Teac falls in that category, as they've had DACs that accept 8x rates for some time.

[wgscott]

I've read the Audirvana documentation.

 

The max filter length adjustment varies directly with the number of filter "taps," i.e., they both describe how many samples the filter acts on. Here is what Miska has to say about number of taps (http://www.computeraudiophile.com/f8-general-forum/we-dont-need-no-stinking-hi-rez-26580/index6.html#post486315):

 

 

 

[Emphasis added.]

 

With more taps (greater the "length" setting in iZotope), one gets "better" frequency operation (meaning deeper cut) and worse time resolution (meaning more ringing).

 

Sorry, I was wrong. It is the steepness setting that is changing as one improves the quality setting on the slider:

 

 

[Jud]

Sorry, I was wrong. It is the steepness setting that is changing as one improves the quality setting on the slider:

 

[ATTACH=CONFIG]25219[/ATTACH]

 

[ATTACH=CONFIG]25220[/ATTACH]

 

Since 0 steepness is no filter at all, I can definitely see how that would be fastest.