Digital Filters, Minimum Phase and Upsampling:
Digital Filters, Minimum Phase and Upsampling:
While I have already mentioned digital filters and upsampling in separate posts, I thought it could be nice to compile some of that in a single post. So here are my findings on the subject: nothing new, just a compilation of thoughts.Digital filters:
One of the discriminating factors between different DACs is the digital filters. Most modern chips include a built-in digital filter (for cost savings reasons) but they perform very poorly in general. But how much difference does it really make?
In my current DAC, I have the possibility to swap between 2 digital filters: the PMD100, which is a HDCD capable digital filter made by Pacific Microsonics, and the DF1704 digital filter made by TI. The two digital filters have 2 distinct sonic signatures (for those interested in the details, they can read my review of the audio-gd dac19mk3 where other users have reported similar differences between the 2 filters).
When playing 16/44 data, the DF1704 is fast and analytical, slightly on the bright side of neutral and the PMD100 is warmer and more “analogue” like.
While the PDM100 is supposed to be limited to 24/55, it seems to work at 24/88 (but not 24/96).
Well, by playing (native) 24/88 data on both, they sounded closer than they did at 16/44. The DF1704 smoothed out and the PMD100 had a little bit more sizzle on top.
That was similar to what I have experienced with some sigma-delta DACs that I have found to sound harsh at 16/44. Usually they sound much better with 24/96 data or upsampling to 24/96. The reason behind the perceived improvement is not due to the fact that we hear the extra data. It is rather due to the fact that we do not have to hear as much the nastiness of the digital filters operating close to our audibility range.
On many digital filters whether they are outboard filters (such as the DF1704) or built-in into the dac chip (such as the CS4398) the designer has sometimes the choice between a sharp/fast roll-off and a slow roll-off.
The sharp roll-off usually gives a flat measured frequency response on the 20-20,000Hz range, which looks good for published measurements. But that flat frequency response up to 20K comes at the expense of phase performance (which has an impact on soundstaging among other things).
The slow roll-off on the other hand is down from 1 to 3 db (in general) at 20 khz but has a better phase linearity, which is more audible to the human ear.
So when given the choice between what looks good on paper (a flat frequency response) and what sounds nice to the ear (phase response), most designers choose the first option.
I have looked at the measurement of many DACs on the stereophile website and some of the most expensive and better sounding DACs seem to have that roll-off on the highs. Weirdly enough most budget DACs and soundcards seem to have a flatter high frequency response. I am not saying that we shouldn’t pursue a flat high frequency response. But I am just saying that when you see a budget DAC that has a flat response for 16/44 data beyond 20 kHz, and if the DAC doesn’t use a fancy DSP, it is most probable that it has sacrificed the phase response for the good looking Frequency response measurements.
By the way, if you are still thinking that a flat frequency response is absolutely necessary to have a realistic sound, just think about the following.
Let’s suppose you listen at a violin in very reverberant room, then you listen to a violin in very damped room of a different size from a different distance. Then listen to a violin in a speaker system that has been equalized to have a flat frequency response. As far as I am concerned, while situation 1 and 2 will have different Frequency Response measurements, my brain will analyze and detect the sound as being real and live. When I will listen to situation 3, I most probably won’t be fooled thinking that the violin is real.
Fortunately, there is a way to go partially past those poor digital filters. By upsampling data to 24/96, you minimize the audible effects of those filters, as any aliasing/distortion they generate will be pushed further up in frequency range.Impulse response & Minimum Phase Filter:
While I have mentioned the Frequency response and phase response of digital filters aspects, I haven’t mentioned yet the impulse response.
When compared to each other, a fast roll-off filter has a lot more pre and post ringing than a slow roll-off filter. That pre-ringing is one of the reasons some people have complained about digital playback and kept using analog sources (Vinyl, tapes…)
So by choosing a fast roll-off over slow roll-off, many DAC and sound card makers are willing to sacrifice the phase and impulse response in order to have a nice looking RMAA graph.
However, while the slow roll-off filter minimizes the pre-ringing, it doesn’t get rid of it entirely. There is a growing number of CD players and DACs that provide a new option which the Minimum phase filter. According to their research, the human ear would be less sensitive to post ringing than to pre-ringing.
For those who are interested, Ayre wrote an interesting paper on Minimal phase filters (here: http://www.ayre.com/pdf/Ayre_MP_White_Paper.pdf
Again those limitations (pre and post response) are far worse at 16/44 than they are at 24/96.
That is once again (in my opinion) the reason why some people hear a lot of improvement when using up-sampling or native 24/96 data. I think that despite some measurements (SNR, THD), not many systems have a true resolution greater than 16 bits. If that were the case (let’s say 20 bit resolution), those systems would be indistinguishable from the reality which is not the case. So, playing back 24/96 data on the DACs can improve the playback experience in comparison to 16/44 simply because most digital filters are terrible at handling 16/44. That is perhaps one of the reasons that we see so many DACs today that use ASRC/upsampling. (But then again those ASRC chips do not seem to sonically sound good and can also benefit from a good upsampling from the source, but that is another story …)SoX upsampling in Foobar:
Among the nice things that come along with using a computer as a source is the ability to try different playback methods and upsampling methods.
Personally, I have settled for Foobar for quite a long time. And recently, I have been using SoX for 2 reasons: It is a relatively transparent upsampler (it doesn’t harm low level details) and it is very customizable.
In my case, I have found the most consistent and enjoyable results with upsampling to 24/96 (weirdly it sounding better than upsampling to 24/88) and with the minimum phase setting. It generally results in a more coherent soundstage that is pushed a little bit further back, and the sound is usually smoother. But those changes vary from one DAC to another, so it is impossible to make generalizations about the effects.
I have also tried different passband settings. The Stock setting is at 95%. When set at 99% the soundstage shrinks and everything gets tighter. When setting the passband at 90%, the soundstage becomes a little bit bigger and the sound and everything sounds a little bit softer and less defined.
Those different settings are a great tool for fine tuning the sound.
While I don’t use the SoX all the time with my reference DAC (which uses a slow roll off DF1704 digital filter), I found that it improved considerably the listening experience with the other DACs I have on hand.