rePhase, a loudspeaker phase linearization, EQ and FIR filtering tool

[pos]

by the way, I forgot some parts of the changelog, so here it is again

0.9.0  2012-10-28
    - new file format '.rephase', saving/loading all settings, including
      correction settings, impulse settings, and graph settings
      (old '.jason' files can still be loaded, but impulse and graph
      settings get reset)
    - impulse file is now a three-part thing: directory, filename, and
      format extension. The directory is the only thing that is not saved
      in the '.rephase' file
    - '.rephase' files can be loaded upon start (as a parameter or by drag
      and drop on rephase.exe) or by drag and drop on the user interface
    - new offset option "float" for fractional sample centering, avoiding
      HF ripples in the impulse when the phase target is not a multiple
      of 180° at the Nyquist frequency
    - make "rectangular" the default window function: this should be the
      best choice for phase-only corrections, and "complex" windowing is
      not needed anymore with the "float" offset
      Note: "rectangular" window is still likely to be the worst choice
      for filter generation, when gain target goes far below 0dB...
    - new "ovelapping" filters, to be used for example in the midbass
      region, under Schroeder's frequency...
    - stereo wav formats are now available
    - bug correction in offset calculation in time=inv mode

[Barleywater]

I used:

Untitled.png

And various combinations of settings; all produced ringing.

Real point is filter mismatch and degree to which it impacts results with convolution engine. Allpass filter generated with Cool Edit convolved with time reversal of itself produces clean Dirac, likewise rephase linearization filter convolved with time reversal of itself produces clean Dirac.

When rephase Linkwitz-Riley 48dB/octave 1kHz linearization filter is convolved with Cool Edit Linkwitz-Riley 48dB/octave 1kHz allpass filter, the slight mismatch appears like this:

(click to enlarge)
convolved mismatch.png

I've seen similar results in past when filters recovered from DCX2496 are convolved with time reversed filters generated with Cool Edit.

On the other hand,when correction filters are generated by mathematically derived inverse transfer function of measured responses, very controlled results are possible.

Regards,

Andrew

[pos]

Hello Andrew,

I was indeed able to reproduce what you got:
I generated a LR48 dB/oct at 1khz with rephase (LR48 in FIR section, and LR48 linearization in time=inv mode) and did the same with cool edit (2 4th order butterworth cascaded on a dirac).
The impulse generated by rephase has some ringing that the one generated by cool edit does not have.
In fact the impulse generated by rephase has more ringing on the left of the impulse, but the impulse generated by cool edit has more ringing of the right side (which is normal I suppose?).

Thank you for bringing that up. This might be due to some 32bit rounding in the FFT engine, I will see how this can be improved.

The impact should be quite limited though: this ringing is very low in level (already much lower than what it was in rephase 0.8.4, thanks to the fractional sample centering) and is at the Nyquist frequency.

[pos]

by the way, version 0.9.1 was uploaded to sourceforge shorly after version 0.9.0 (~13 hours ago), because some bugs had been introduced:

0.9.1  2012-10-29
    - centering 'int' option was not working, this is now fixed
    - more explicit error message when loading a wrong setting file

namely, the 'int' centering behavior (emulating the centering of rephase 0.8.4) was not working, and impulse offset was always a fractional sample...

[Barleywater]

......cool edit has more ringing of the right side (which is normal I suppose?).

Cool Edit Pro2.1 running on XP SP2: LR 48dB/octave generated starting with Dirac produces all zero to left of causal Dirac, when waveform is reversed in time all zeroes appear to right. I start with stereo Dirac, apply HP Butterworth 4th order twice to left track, and apply LP Butterworth 4th order twice to right track, both tracks are then summed producing all-pass result with non zero values spanning 2101 samples (48kHz sample rate):

Normal time order:

left of causal.png

Final sample:

last causal sample.png

Convolution of the 2101 sample filter with a time reversed copy yields:

dirac recovered.png

Convolution is Aurora plugin, applied with 0 scaling or DC correction. Recovered Dicac has normalized value of one. Total math has been done on 6092 samples as apparent in above picture.

Andrew

[pos]

rePhase 0.9.2 is online

Download: rePhase on sourceforge

0.9.2  2012-11-04
    - added back '24bit LPCM mono' output format, missing since 0.9.0
    - improved up/down key bindings on frequency entries
    - added up/down key bindings for taps entry, with color warnings for
      extreme values
    - new '1st order' and '2nd order' linear-phase filters, meant to be
      combined with an existing (and already corrected in phase) rolloff
      to obtain a linear-phase acoustical Linkwitz-Riley filter
    - made sure 'middle'+'float' centering ends up within -0.5/+0.5 sample
      from middle (was -1/+0.5)

[peterhenk]

Hi Pos,

Most probably a very stupid question..... So be kind.

Can I use this to get linear phase improvements in my DCX software? If yes, how?

Originally Posted by pos

You can use rePhase to linearize passive loudspeakers, active ones, and turn IIR active crossovers into linear-phase crossovers (turn a DCX2496 into a DEQX or a Dolby Lake ).
[/code]

Thank you
Peter

[pos]

Hi Peter,

Not a stupid question at all, and in fact I am really sorry I did not write any documentation for this software

First, the comparison between the DCX and the Dolby Lake was a bit of a joke: you *can* obtain a linear phase filter (and acoustical response) with a DCX and rephase, but you will never get the audio quality (better DACs, better filtering algorithms, etc..) and power (brickwall filters, arbitrary slopes and eq, ...) of a Lake with a DCX.

Still, you will indeed be able to linearize the phase shifts introduced by the IIR filters in the DCX (and drivers/box) with rephase, and get good results (especially if you have crossover points under 1khz, which are more sensitive to phase shifts).
To do so you will only need to dial the filters you are using in your DCX in the "filter linearization" tab in rePhase, and generate an impulse for these (ideally you will want to dial your acoustical filters, not only the electrical ones...).
This is similar to phase arbitrator (which happen to have a user manual )

Then in order to use the generated convolution (carrying your phase correction), you will need to find a way to introduce a convolution engine in you audio chain, before the DCX.
If your source is a PC (or Mac) then it should be easy enough: there is a lot a convolution plugin that you can use, depending on your favorite audio player).
If you do not use a PC things will be more difficult: you will have to introduce a piece of hardware to do the convolution.
The minidsp openDRC is currently the only device (that I am aware of) that can do such convolutions.

[peterhenk]

Thank you for the reply,

I am using a mac as source, and a heavily modified DCX. So I will be playing around this Xmass with your software!!

Merci
Peter

[Pano]

Wow, how did I miss this before? Will try this for sure. Thanks for posting.