MEASUREMENTS: HiFiBerry DAC+ Pro, PCM5122 digital filters & further thoughts on filtering...

A summary slide from this presentation slide set. Just a reminder of the general desirability of digital filters providing higher accuracy compared to analogue...

Thanks to Jim Ambras in the discussion of the measurements for the Raspberry Pi 3 & HiFiBerry DAC+ Pro for pointing me to some settings available! So in the "Wait! There's more..." department, there are some "advanced" settings available in Alsamixer for the DAC+ Pro I2S HAT board you might wish to experiment with. Let us have a look at the different digital filters built into this very affordable DAC and how it changes the sound objectively...

The settings are easy to get at and changes can be made "on the fly" so you can quickly hear the difference (or not) that the digital filter makes. A great way to explore for yourself what your preferences are... So, if you have Volumio installed, all you do is get into an SSH session (for example, use PuTTY in Windows) into the command line of the Pi (remember for Volumio, login and password are both "volumio"), then issue the "alsamixer" command. You should see this:

Then press the F6 key to select the DAC (DAC+ Pro selected in this case of course):

After pressing enter, you'll see the digital filter setting (aka "DSP Program") as the first item which you can change with the UP/DOWN arrow key on your keyboard:

As you can see, I have the "Low latency IIR with de-emphasis" selected as the "DSP Program". There are 5 selections you can make:

- FIR interpolation with de-emphasis [DEFAULT] ("Normal" 8x oversampling)
- Low latency IIR with de-emphasis
- High attenuation with de-emphasis
- Fixed process flow [no sound output when I choose this one]
- Ringing-less low latency FIR (I believe this is called "Asymmetric FIR (8x)" in datasheet)

Since I could not get any sound output with the "Fixed process flow" setting (I wonder if it's a programmable setting which is unused in this case), there are 4 settings to have a look at here. Let's plot out the characteristics of these digital filters which might help you think about what you're listening to and maybe even decide how much of a difference you think this all makes!

I. Impulse Response

As you can see, we have a variety of impulse responses characteristic of the filters to choose from. These are easily differentiated on these time-domain plots... By looking at the morphology, we can predict the effect. The default "FIR interpolation" filter is the standard linear phase moderately steep setting previously measured. "Low latency IIR" is a minimum phase, steep filter. "High attenuation" is linear phase, but with more taps for a sharper/steeper filter but more pre- and post-ringing. Finally, we have the "Ringing-less" setting which is a combination minimum phase with early roll-off, similar to the PonoPlayer but less extreme early roll-off of high frequencies and concomitantly a little more of the post-ringing.

II. Digital Filter Composite

As I have done over the last year, these are the "digital filter composite" graphs I've been using to demonstrate the accuracy/strength of the digital filters to suppress imaging distortions and intermodulation distortion. Again, see this Stereophile discussion with Juergen Reis for original ideas around this type of testing.

These first two graphs actually look very similar despite the obvious time domain differences above in the impulse response plots. As filters in the frequency domain, these are essentially equivalent, both moderately steep in nature with generally good suppression of imaging, intermodulation and harmonic products. There's at least 60dB suppression of distortion above Nyquist.

As the name implies, this is a stronger attenuation filter compared to the above two. It's steeper, has more potent suppression of harmonics overall. No surprise then that the time domain impulse response consists of stronger "ringing" as well.

Finally, we have the "ringing-less" filter:

Yup. There is a price to pay for minimizing pre- and post-ringing! As you can see, with a poor reconstruction filter, a lot more distortion products are allowed to pass through beyond the Nyquist frequency. As predicted with the reduction of ringing, the white noise (yellow) tracing shows an earlier and more gradual roll-off.

III. RightMark results

Realize that the digital filter effect is most significant with the CD-resolution 44kHz sampling rate since Nyquist is closest to the range of human hearing. I was curious to see if the different settings changed the distortion characteristics, noise, and dynamic range measurements. Here's the summary:

Remember that RightMark is calculating the overall results from 20-20kHz. Overall, we see no evidence of the filters being able to cause big effects. Remember that the effect of the digital filters are essentially in the high frequencies. Since hearing perception is logarithmic, these digital filter effects should be minimal and consistent with limitations on human auditory acuity.

As for specific graphs:

Frequency Response (zoomed in from 4kHz to highlight differences)

Noise level

IMD+N sweep

You can see that I zoomed into the frequency response graph in order to better appreciate the earlier roll-off from the "Ringing-less" filter setting. About -1.3dB down in frequency response by 20kHz compared to -0.5dB for the others. Not a big deal nor easily audible (compared to the PonoPlayer which is closer to -4dB at 20kHz) unless one is young with truly superb hearing (sorry gentlemen over the early 20's, we generally will have to concede in this regard).

Noise floor is the same with all settings - not unexpected.

The IMD+N sweep results basically shows the earlier roll-off of the "Ringing-less" filter as appearing like higher distortion.

IV. Concluding Remarks...

I'm not going to spill too many characters on this other than to say that now you have knowledge of what the PCM5122's filters are doing, go listen for yourself to contextualize the subjective perception. When it comes to truly understanding, this synthesis of objectivity and subjectivity is I believe what should happen, not just as audiophiles but in other domains of life as well where applicable. As I posted previously about antialiasing digital filters, my personal sense is that the effect is subtle. It's great in situations like this to have the ability to switch the setting quite easily, literally "on the fly" for quick A/B comparison listening.

One interesting comment Jim Ambras made in his response was this: "The few people who have played around (with) this seem to prefer the Ringing-less low latency FIR as do I." As you can see from the results, this is the weakest of the antialiasing settings and with strong high frequency signals, will allow quite a bit of ultrasonic distortion to pass through. Notice there is some euphemism in the wording of the filter name here biasing the time domain characteristics as if less ringing is a good thing when in fact, as we've discussed before, this will have a negative influence on the frequency domain. So "Ringing-less" in this case is also the same as "Phase-shift-more" (minimum phase effect) and "Aliasing-distortion-more" (weak filter). No free lunch...

I think it would be very interesting to see the results of a blind test between having no anti-imaging filter (ie. NOS DAC sound) vs. a weak filter (like this "Ringing-less" setting or PonoPlayer) vs. something reasonably strong like the default ("FIR interpolation"). Technically, having at least a moderately strong anti-imaging filter is good for fidelity since it removes the ultrasonic products that should not be there in the analogue output. However, people will have preferences and I can appreciate companies like Ayre who provide both "measure" (stronger) and "listen" (weaker) filters to accommodate both technical accuracy/fidelity as well as personal preference when listening.

Realize that even in the audiophile community, products will vary and there will be companies that actually vouche for a stronger antialiasing filter with more ringing. For example, check out Chord and the hype around their FPGA filtering; often making a point of telling us how many taps their FIR filter utilizes (26,368 for the ~US$4-5k Hugo TT for example). Just have a look at the impulse response measurements and the length of that linear pre/post-ringing with a very steep "brick wall" filter. Although I haven't seen measurements for the Chord DAVE, it is said to use 164,000 taps with an asking price of US$13k without a stand... Impressed by big numbers anyone?! Anyhow, this is all to say that in reality, one could go both ways with these filters in the audiophile world and still charge lots of cash and nobody consistently says one is clearly better than another... In other words... Subtle at best.

-------------------------------

Well everyone, hope you're all enjoying October. I'll actually be out of the country again for some work-related duties for a few weeks...

But before I go, let me show you what I just borrowed for a spin:

Well, well, well, what have we got here??? Might or might not be interesting depending on your perspective! :-)

Enjoyed a few albums this week including Rachel Barton Pine's Joachim, Brahms: Violin Concertos (2CD's, 2003, ~DR16) and Empire of the Sun's Ice On The Dune (2013, DR5 - boo). Musical surprise of the week while relaxing on some pop - enjoyed the new Lady Gaga album Joanne (2016, cheapo $3.99 Amazon MP3 here - I'm assuming they're doing this as a virtual loss leader, back in 2011, the deal was better at $0.99 with Gaga). The songs "Sinner's Prayer" and "Just Another Day" were catchy. Some of the songs have an '80s vibe (eg. "Come To Mama" even has a sax part, and parts of "Hey Girl" sounds like pop funk from the late 70's). Modern "shouty" dynamic compression of DR7 not unexpectedly.

Until next time... Enjoy the music everyone.