• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Minimum Phase vs Linear Phase

OP
Ron Texas

Ron Texas

Master Contributor
Joined
Jun 10, 2018
Messages
6,075
Likes
8,908
I believe the default filter setting on the popular Topping DX3 Pro is short delay, sharp. When the Grace M900 was tested here there was significant distortion added by the short delay slow filter. The latter may have been a bad implementation.
 

maty

Major Contributor
Joined
Dec 12, 2017
Messages
4,596
Likes
3,160
Location
Tarragona (Spain)
[PDF] http://download.oppodigital.com/UDP20X/UDP-205_User_Manual_English_V1.0.0.pdf

Oppo-UPD-205-filters.png


Again, minimum phase. It has ESS ES9038PRO.
 

maty

Major Contributor
Joined
Dec 12, 2017
Messages
4,596
Likes
3,160
Location
Tarragona (Spain)
No need to search anymore. It must always be by default, regardless of the chip, minimum phase. Maybe, if there are internal oversampling, linear phase. But the traditional DAC work with MP by default since ever.

This is how the MP3 vs FLAC and ... the years go by and many keep turning to overcoming issues. Even in forums where there is supposed to be a qualified person.
 
Last edited:

daftcombo

Major Contributor
Forum Donor
Joined
Feb 5, 2019
Messages
3,687
Likes
4,068
Once and for all, the only advantage of minimum phase on linear phase is no latency when watching a movie.
 

mansr

Major Contributor
Joined
Oct 5, 2018
Messages
4,685
Likes
10,700
Location
Hampshire
Once and for all, the only advantage of minimum phase on linear phase is no latency when watching a movie.
Wrong. The filter delay is only a small part of the total audio processing latency which always needs to be taken into account, along with the video latency, in order to maintain synchronisation. Besides, a shift of less than 10 ms or so isn't noticed anyway.

Where minimum phase filters are often necessary is in live effect processors. Playing an instrument or singing becomes impossible when what you hear is delayed by more than a few milliseconds.
 

MRC01

Major Contributor
Joined
Feb 5, 2019
Messages
3,422
Likes
4,029
Location
Pacific Northwest
The 'fun' part is that you are showing signals that don't exist in any music files. All 44.1 recordings should have passed through a steep brickwall filter on the encoder side (when down-sampled properly from the recording format) so frequencies that are reaching Nyquist should not be there. Square-waves that were recorded through the analog input or are down-sampled won't have such steep rising and falling edges.
Besides, the upper treble 'energy level' in recordings (almost) never reaches near FS. So stuff that 'mirrors' back due to slow filtering also will be lower in amplitude (unless one is daft enough to use filterless at 44.1/48)
That's why I suggested to use (recorded) music and nulling software if you want to find out where the audible differences are.
...
As I mentioned earlier, with most music I can't hear the difference between these filters. But there are a few tracks where I can. The 500 Hz square wave is one of them. What's interesting is the perception of the difference is the opposite of what I expected: on the slow filter, this signal sounds like it has just a touch more "edge" or "buzz", similar to hearing a pure midrange sine wave that has a barely perceptible (say 1%) distortion. Yet the FFT (and the FR curve) shows the reverse, the sharp filter having slightly more HF energy. Perhaps what I'm hearing is a phase shift that is accentuating certain frequencies, which wouldn't show up in FR or FFT.

Also, the step responses from REW are more similar than the actual square waves. They don't show the slow filter's over-shoot and longer ringing. This makes me wonder if deriving the impulse response from the sweep (which REW does), while mathematically exact, is less than exact in the real world. So you get a mathematically predicted step response that doesn't necessarily match actual step response. Maybe there's a different reason they don't match.

The square wave response here shows that the slow filter on my DAC is minimum phase. But not all "slow" filters are minimum phase. Some are still linear phase, but use a wider transition band (like filters #1 and #5 on the WM8741). This leaves me wondering what that would look like in measurements. From a math & engineering perspective, that would seem to be the ideal filter.

Anyway, I normally listen to music, not test signals. On those few musical tracks where the difference can be heard (such as the Drums & Bells snippet I posted earlier), it's so subtle I honestly can't say which is more realistic or express any preference. So the experiment shows the switch actually does something, but it doesn't make any difference to me in actual musical listening to music. So the experiment was educational, but I'll leave it in the sharp position simply for peace of mind knowing it's more correct.
 

solderdude

Grand Contributor
Joined
Jul 21, 2018
Messages
15,891
Likes
35,912
Location
The Neitherlands
From a math & engineering perspective

From that perspective the fast linear phase filter is closest to the ideal.
There is nothing ideal about a slow filter... that is it only SEEMS more ideal for illegal test signals that do not exist in any music file.
Why bother to get that reproduced perfectly and allow mirror images and HF non harmonic inaudible garbage that might (doesn't have to) create audible artifacts in the chain behind the DAC.

Slow filters are not because A: they roll-off too soon B: the mirror images are too high in level and could become audible when there are nonĺinearities in the chain after the DAC.

When you would want to listen to a 500Hz squarewave you should simulate one that has gone through a steep filter cutting off everything above 20kHz and play back that file.
This could actually be in a file. a 500Hz digital squarewave cannot possibly exist in a music signal.
When listening to that you could hear effects that won't ever be there IRL.

Indeed I would say just listen to music. Start of with the most 'ideal' filter, experiment with the filters and if you prefer one that works good on your system (may not be so on someone else's) and is better suited to your taste use that.
Just realize that filter is not 'better' it merely sounds better to the one preferring it.
 
OP
Ron Texas

Ron Texas

Master Contributor
Joined
Jun 10, 2018
Messages
6,075
Likes
8,908
From that perspective the fast linear phase filter is closest to the ideal.

@mansr wrote something like that a few pages back. None the less I prefer (not blind tested) the short delay(minimum phase), fast filter for listening and several products use it as their default including an expensive Bryston DAC where the filter is fixed. It's entirely possible that a large portion of the population would prefer the linear/fast delay filter for listening.
 

MRC01

Major Contributor
Joined
Feb 5, 2019
Messages
3,422
Likes
4,029
Location
Pacific Northwest
From that perspective the fast linear phase filter is closest to the ideal.
There is nothing ideal about a slow filter... that is it only SEEMS more ideal for illegal test signals that do not exist in any music file.
Why bother to get that reproduced perfectly and allow mirror images and HF non harmonic inaudible garbage that might (doesn't have to) create audible artifacts in the chain behind the DAC.
...
There are many different kinds of slow filters and it sounds like you and I mean something different by this term. The ideal slow filter would be the same as the "sharp" filter, linear phase, except that it would use the widest possible transition band. For example, consider the WM8741 (which I think is a typical delta-sigma DAC chip) and suppose you're listening to a track sampled at 96 kHz. The standard sharp filter is flat to 0.4fs which is 38.4 kHz which makes a transition band (48 - 38.4 = 9.6 kHz wide). The slow filter is flat to about 0.204fs which is 19.5 kHz which makes a transition band (48 - 19.5 = 28.5 kHz wide). Both are fully attenuated at or before 0.5fs, so neither has any aliasing of HF noise. The only difference between them is the slow filter has less passband ripple, which should be closer to the ideal W-S response. Essentially, the sharp filter uses the extra bandwidth to reproduce inaudible supersonic sounds, while the slow filter uses it to have a more gradual slope that is more transparent in the passband.

At CD quality, the difference is moot. The transition band is already so narrow there can't be any such thing as a slow filter, unless it attenuates the passband, or doesn't attenuate fully at Nyquist. Neither of which is acceptable.
 
OP
Ron Texas

Ron Texas

Master Contributor
Joined
Jun 10, 2018
Messages
6,075
Likes
8,908
At CD quality, the difference is moot. The transition band is already so narrow there can't be any such thing as a slow filter, unless it attenuates the passband, or doesn't attenuate fully at Nyquist. Neither of which is acceptable.

I believe at Redbook quality slow rolloff filters are attenuating the pass band some. Older listeners can't hear it, though.
 

solderdude

Grand Contributor
Joined
Jul 21, 2018
Messages
15,891
Likes
35,912
Location
The Neitherlands
When it comes to audible effects of reconstruction filters I always mean 44.1 and 48kHz files.
Most people have 44.1kHz music anyway.
For 88.2kHz and up the steepness of the filter is far less important as the roll-off is well above the audible band and mirror images are well above 44kHz so will not be reproduced by the majority of speakers/headphones anyway.

I have no beef with preference at all. Many folks prefer something.
The thing I usually object to is that the persons preference equals 'better'.
It is not, the person just prefers it.
Regardless if it is tested blind or sighted.

I don't think any person can detect a passband ripple smaller than 0.5dB, certainly not in high frequencies.
 
OP
Ron Texas

Ron Texas

Master Contributor
Joined
Jun 10, 2018
Messages
6,075
Likes
8,908
I have no beef with preference at all. Many folks prefer something.
The thing I usually object to is that the persons preference equals 'better'.
It is not, the person just prefers it.
Regardless if it is tested blind or sighted.

That's a reasonable approach. In the research done at Harmann results were tabulated as listener preferences. These preferences were statistically significant and consistent across different groups of listeners. However, there is a sizable minority with regard to their preferences. It's why panel speakers which did poorly in Harman's labs continue to sell and satisfy their owners, some of whom are members here.
 

RichB

Major Contributor
Forum Donor
Joined
May 24, 2019
Messages
1,946
Likes
2,611
Location
Massachusetts

daftcombo

Major Contributor
Forum Donor
Joined
Feb 5, 2019
Messages
3,687
Likes
4,068
Wrong. The filter delay is only a small part of the total audio processing latency which always needs to be taken into account, along with the video latency, in order to maintain synchronisation. Besides, a shift of less than 10 ms or so isn't noticed anyway.

Where minimum phase filters are often necessary is in live effect processors. Playing an instrument or singing becomes impossible when what you hear is delayed by more than a few milliseconds.

You are backing what I said: only advantage is low latency.
 

maty

Major Contributor
Joined
Dec 12, 2017
Messages
4,596
Likes
3,160
Location
Tarragona (Spain)
I was reviewing the Okto review and ...

https://www.audiosciencereview.com/...d-measurements-of-okto-dac8-8ch-dac-amp.7064/

And finally the ESS Sabre DAC settings which I left alone:
index.php


http://archimago.blogspot.com/2017/...howComment=1500132235162#c8640249355692475027

[ The User's Manual for the Mytek Manhattan II has various options for PCM and DSD filters...

8.2.11 PCM FILTER shapes

FRMP - fast roll-off, minimum phase filter
• SRMP - slow roll-off, minimum phase filter
• FRLP - fast roll-off, linear phase filter
• SRLP - slow roll-off, linear phase filter
• APDZ - apodizing, fast roll-off, linear phase filter
• HBRD - hybrid, fast roll-off, minimum phase filter
• BRCK - brickwall filter


8.2.12 DSD FILTER (DSD Filter Bandwidth)

• AUTO – the filter is selected automatically depending on DSD rate: for DSDx64 - LO, DSDx128 - MED, x256 – HI. It is highly recommended to leave this option enabled.
• LO - 47,44 kHZ IIR Filter
• MED - 60kHz IIR Filter
• HI - 70 kHz IIR Filter ]

It would be nice to ask the Okto designer about the filters and in what situation is one better than another (each Okto filters).

BTW, the image says: FIR filter FRMP, should not it be IIR filter?
 
Last edited:

MRC01

Major Contributor
Joined
Feb 5, 2019
Messages
3,422
Likes
4,029
Location
Pacific Northwest
I believe at Redbook quality slow rolloff filters are attenuating the pass band some. Older listeners can't hear it, though.
That's true. The passband attenuation from the slow filter in my DAC is flat to 18 kHz and about -3 dB @ 20 kHz. I certainly can't hear that, even most young people can't. Of course that doesn't mean it's transparent; it has other effects that may be audible especially when listening to certain kinds of sounds.

BTW, there's another interesting aspect of the sharp filter in my DAC (a WM8741 in max OSR mode with filter #3). Its square wave is not symmetric. It rings more after an impulse (on the L side of a peak), than before an impulse (on the R side of a peak). A pic's worth 1,000 words:
1568497324532.png


The slow filter exaggerates this further:
1568497426450.png


The ideal W-S reconstruction would be perfectly symmetric. So even the sharp filter doesn't match the ideal response. However, since this is not the direct analog output, but from a digital recorder, I don't know whether this asymmetry came from the DAC's D-A, or the recorder's A-D.
 

NTK

Major Contributor
Forum Donor
Joined
Aug 11, 2019
Messages
2,656
Likes
5,819
Location
US East
I just came across an article at EDN. A commenter raised an interesting point regarding minimum phase filter — the phase distortion from them often result in an increase in the peak voltages over the "original" unfiltered signal.

I did a simulation myself to compare the sinc filter with the most common types of analog equivalent IIR filters. I used a square wave as the original signal. All of the IIR filters has higher ripple peaks than the sinc filter. The higher ripple peaks will mean that the minimum phase filters are more prone to result in clipping or overloading the output circuitry.

I wonder if this (the amount these ripples eat into the headroom of the DAC) is something ASR should investigate.

Figure.png
 

SIY

Grand Contributor
Technical Expert
Joined
Apr 6, 2018
Messages
10,383
Likes
24,749
Location
Alfred, NY
I just came across an article at EDN. A commenter raised an interesting point regarding minimum phase filter — the phase distortion from them often result in an increase in the peak voltages over the "original" unfiltered signal.

I did a simulation myself to compare the sinc filter with the most common types of analog equivalent IIR filters. I used a square wave as the original signal. All of the IIR filters has higher ripple peaks than the sinc filter. The higher ripple peaks will mean that the minimum phase filters are more prone to result in clipping or overloading the output circuitry.

I wonder if this (the amount these ripples eat into the headroom of the DAC) is something ASR should investigate.

View attachment 33374

Did you bandwidth limit the square wave? Otherwise it's an illegal signal.
 

NTK

Major Contributor
Forum Donor
Joined
Aug 11, 2019
Messages
2,656
Likes
5,819
Location
US East
Did you bandwidth limit the square wave? Otherwise it's an illegal signal.
No I did not. The only thing it will do is to make the sinc filter look better. If you use a band limited square wave, there will be no effect on the sinc filter. For the minimum phase filters, it certainly will have an effect (may be very small).

The reason I used a square wave if that the impulse (which we almost exclusively talk about here) is also not a band limited signal. So I am not doing something too different. Guess what. The only band limited signal that will give an "impulse looking" digital signal train is the sinc function. The sinc filter just perfectly reproduces that! Just like what the digital sampling theorem says.

sinc.JPG


[Edit, added this paragraph] BTW. If the digital pulse train fed into the DAC "looks" like the square wave. Those signal traces will be the results if we use those filters for reconstruction. Thus I believe my simulation was legit.
 
Last edited:

MRC01

Major Contributor
Joined
Feb 5, 2019
Messages
3,422
Likes
4,029
Location
Pacific Northwest
You can see this happening in my examples too. The min phase filter overshoots the step by a bigger margin. In my case, about 3 dB more. The ripple is close to Nyquist so clipping is not a practical concern with music.

And SIY makes a good point: an analog signal that is properly BW limited below Nyquist, when sampled can't jump from 0 to full scale in a single sample. But it could jump to some other value, like -12 dB. Even so, square waves don't exist in nature since a infinite rate of change in air pressure is impossible. But it can change faster than Nyquist, which must be removed before encoding. Removing it -- a low pass filter -- will create the ripple if its slope is steep enough.

It's worth noting that the Gibbs effect ripples are not a digital artifact. They're the natural effect of a low pass filter with steep slope, whether digital or analog.
 
Top Bottom