Filter brewing for the Soekris R2R

Disabled Account
Joined 2005
Move comparative tests

The second test is a 19K test tone. Again no calibration, this was just set to the same level as the white noise.

Same order of appearance as above.

MB1 is a bit cleaner in the audio band than Søren filter. The MPSRO shows the obvious image at around 24kHz which is not present in filters with steep roll-off for anti-imaging. It's also has a lot more junk in the audio band. It's worth noting that the worst case additional junk is around -80dB on the test tone.
 

Attachments

  • MB1_19K.png
    MB1_19K.png
    222.9 KB · Views: 992
  • 1021Default_19K.png
    1021Default_19K.png
    183.2 KB · Views: 944
  • MPSRO441v2_19K.png
    MPSRO441v2_19K.png
    198.8 KB · Views: 896
Last edited:

TNT

Member
Joined 2003
Paid Member
spzzzzkt is right on the money.

These 3 filters are all steep. They are also symmetrical i.e. they have equal pre and post "ring" around a pulse.

MB1=nuttall - data1
MB2=Taylor - data2
MB3=Cheby(chev) - data3

In the pulse plot, all three are actually in there in one and the same plot.

Group delay over frequency (derivate of phase) is constant for all three filteters - only one shown. These should have excellent time coherence or as usually expressed "phase correct".

This is what I have but I (get) answers on any questions.

btw - don't forget to pull the serial connector when finished loading and starting to listen - you don't want your computer earth in your D/A.

I liked MB2 a lot. Had the best experience so far with this filter. I think MB1 and MB2 sound considerable cleaner than the what I had listen to earlier. The soft roll off filters has a trail of quite nasty hf that is not music but maybe clarity. MB2 was to me more mechanical somehow - not as much music. MB2 made my muscles around the ear relax.

Yes - tonality is a game of matching components but cleanness and coherence can't be. Many of the latest trend tall/narrow speakers do suffer from lack of energy in the warmth/chest band du to less coupling to the room. More baffle area cures this but also some EQ would.

I still feel that the D/A is not as clean in the upper regions as I would have wished - at times it has a little grey hash. See if the unbuffered output provides a solution. Also for these filters I need to hear them over loudspeakers - hopefully this week.

Thanks to Mikael B for designing these filters which he says are all text book filters with a little tweak with windowing.

Filter thumbnails left to right - MB1, MB2 and MB3
//
 

Attachments

  • grpdelay.png
    grpdelay.png
    11.3 KB · Views: 464
  • IR 1_3.png
    IR 1_3.png
    19.5 KB · Views: 466
  • ChebyWin.png
    ChebyWin.png
    17.1 KB · Views: 435
  • taylor.png
    taylor.png
    18.1 KB · Views: 461
  • nuttall_1016_8.png
    nuttall_1016_8.png
    17.8 KB · Views: 918
Last edited:

TNT

Member
Joined 2003
Paid Member
If you experience loss of output level, one possible reason could be that you did not change the "numbercoefficients" parameter implying that not the whole filter is executing. Sören to confirm that this would be the consequence...

-- Format:
-- signature, samplerate, interpolationrate, type, numbercoefficients, multiplier

//
 

TNT

Member
Joined 2003
Paid Member
spzzzzkt is right on the money.

These 3 filters are all steep. They are also symmetrical i.e. they have equal pre and post "ring" around a pulse.

MB1=nuttall - data1
MB2=Taylor - data2
MB3=Cheby(chev) - data3

In the pulse plot, all three are actually in there in one and the same plot.

Group delay over frequency (derivate of phase) is constant for all three filteters - only one shown. These should have excellent time coherence or as usually expressed "phase correct".

This is what I have but I (get) answers on any questions.

btw - don't forget to pull the serial connector when finished loading and starting to listen - you don't want your computer earth in your D/A.

I liked MB2 a lot. Had the best experience so far with this filter. I think MB1 and MB2 sound considerable cleaner than the what I had listen to earlier. The soft roll off filters has a trail of quite nasty hf that is not music but maybe clarity. MB2 was to me more mechanical somehow - not as much music. MB2 made my muscles around the ear relax.

Yes - tonality is a game of matching components but cleanness and coherence can't be. Many of the latest trend tall/narrow speakers do suffer from lack of energy in the warmth/chest band du to less coupling to the room. More baffle area cures this but also some EQ would.

I still feel that the D/A is not as clean in the upper regions as I would have wished - at times it has a little grey hash. See if the unbuffered output provides a solution. Also for these filters I need to hear them over loudspeakers - hopefully this week.

Thanks to Mikael B for designing these filters which he says are all text book filters with a little tweak with windowing.

Filter thumbnails left to right - MB1, MB2 and MB3
//

Should have read: MB3 was to me more mechanical somehow

//
 
Ok so those filters are basically brickwall filters (so called "windowed sinc"), using 3 different windowing algorithm.
This is the simplest possible LP filter design you can do with FIR.
The final shape (steepness and ripples "floor") is dictated by the chosen window.
Similar design to what Soren originally did use, but with more taps which allows for steeper slopes with the same ripples floor...

Attached is an example you can play with (choose different windowing algorithms and generate to see the resulting curve, and also maybe change the filter frequency).
Those that work well in this context are Nuttall, Blackman-Nuttall, and Blackman-Harris.
I don't have Taylor and Chebychev windows in rephase so cannot try these, but I can add them for the next version.

Funny thing is that people wanted a shallower slope and no preringing, but now the steepest linear-phase (ie tons of preringing) filter is preferred ;)
 

Attachments

  • DAC 44,1kHz brickwall.zip
    1.1 KB · Views: 208
Last edited:
Disabled Account
Joined 2005
Ok so those filters are basically brickwall filters (so called "windowed sinc"), using 3 different windowing algorithm.
This is the simplest possible LP filter design you can do with FIR.
The final shape (steepness and ripples "floor") is dictated by the chosen window.
Similar design to what Soren originally did use, but with more taps which allows for steeper slopes with the same ripples floor...

Attached is an example you can play with (choose different windowing algorithms and generate to see the resulting curve, and also maybe change the filter frequency).
Those that work well in this context are Nuttall, Blackman-Nuttall, and Blackman-Harris.
I don't have Taylor and Chebychev windows in rephase so cannot try these, but I can add them for the next version.

Funny thing is that people wanted a shallower slope and no preringing, but now the steepest linear-phase (ie tons of preringing) filter is preferred ;)

Thanks pos!

Of course measuring and listening are two different things ;) It was partly to see what was happening on the outputs and how close it matched the filter curves.

I still like the slow roll-off / little ringing better for listening.

cheers
Paul
 
Disabled Account
Joined 2005
If you experience loss of output level, one possible reason could be that you did not change the "numbercoefficients" parameter implying that not the whole filter is executing. Sören to confirm that this would be the consequence...

-- Format:
-- signature, samplerate, interpolationrate, type, numbercoefficients, multiplier

//

All the filters are done at 1015 with the following header...

Code:
dam1021,44100,8,1,1015,8
Input FIR, 44.1 Khz Samplerate, 0-19 Khz +- 1.83 db, 24 Khz -7.84db

It's more likely related to the slow roll-off in the HF.
The white noise was fairly close, but the 19kHz tone was down, and this is where the slow roll-off is starting to bite.
Checked the response of the filter - it's -3dB at 19kHz.
 
Last edited:
Member
Joined 2009
Paid Member
If you experience loss of output level, one possible reason could be that you did not change the "numbercoefficients" parameter implying that not the whole filter is executing. Sören to confirm that this would be the consequence...

-- Format:
-- signature, samplerate, interpolationrate, type, numbercoefficients, multiplier

//

As I said, there isn't much error checking on filter tools, parameters have to match or consequences varies between subtle differences, to strange sound to no sound....
 
Yeap it sounds like that when I read him commercial website... sure you have a better understanding than me as I don't understand absolutly nothing to those digital filters :D.

The tube filter has a great sucess, certainly adding a descent repartition between harmonics with odds ones : so numeric re tailored analogic somewhere at the output !... with my simple words !:eek:
 
With an active stage following the DAC incorporating a low pass filter seems it could be useful to remove that capacitor on the output of the DAC.

Could not possibly be one of the better dielectrics. Especially if one is duplicating the job with a subsequent circuit!

Either a capacitor based filter or transformer coupling of the line stage should be able to make a much better filter than what is given.