I Have A Theory About Why I Like Linear Phase Filters

[Jud]

I tried a little (yes, blind) experiment yesterday, and it may have borne out some speculation on my part about why I prefer linear phase filters in my system.

 

Using Audiophile Inventory (AuI), I converted the 24/192 version of "Beryl," from Mark Knopfler's latest album "Tracker," to DSD128. I did this twice, once using AuI's linear phase setting, then again using the minimum phase setting. I placed the two in a single location (a RAMdisk) and played both in sequence using Audirvana+. I did not know other than listening which file was which until I confirmed by looking at the metadata afterward (I'd renamed the file that was converted using minimum phase to "Berylmin").

 

The first file in the sequence - which turned out to be the one converted with the minimum phase filter setting - did sound as if it had slightly better, quicker instrumental "attacks" on drums and guitar strings. The second, though not quite as good in that area, did better with imaging - overall soundstage and instrumental placement/localization.

 

I have Vandersteen speakers, which are "time-aligned," i.e., supposedly designed so that sound from each of the drivers arrives at the listener's ear at approximately the same time. The crossover is also designed to preserve phase. So in that context, think about what a minimum phase filter does: (1) It is a "dispersive" filter, meaning the time it takes for a signal to pass through the filter differs by frequency; (2) It doesn't preserve phase. It seems to me, then, that a minimum phase filter should counteract to some extent the soundstage and localization advantages conferred by the design of the Vandersteens, and that is indeed what I heard.

 

The sound from the file that turned out to be linear phase was not perfect, because the attacks were slightly "rounded" in comparison to the more realistic ones from the minimum phase file. But overall, the sound from the linear phase file was more like real players and vocalists located in specific places in a well defined space.

 

Any thoughts from people who know something about filters or speakers, or who just have an opinion about all of this?

[audiventory]

Hi Jud,

 

Thank you for testing minimal phase filter.

 

I applied it several years ago, but while almost nobody interest it.

 

Currently all minimal phase filters is 2G (second generation).

 

Linear phase filters for sample rates D64(2.8 MHz)/128(5.6 MHz) /256(11.2 MHz)/512(22.5 MHz) and 44.1 kHz (after yesterday) replaced to 3G.

 

3G filters technically has less noise/distortions, sometime faster.

 

Technically applying 3G filters to minimal phase filter is possible.

 

I while don’t touched min. phase filters due a few interest.

 

However me seem filters without pre-ringing deserve attention.

 

Me very interesting discussion about this almost unexplored part of audio conversion.

 

Best regards,

Yuri

[JohnSwenson]

I personally prefer intermediate phase filters, you can split the difference between what you heard, they have some pre-ringing, but less than linear phase and less post ringing than minimum phase.

 

John S.

[rodrigaj]

Hi Jud,

 

Thank you for testing minimal phase filter.

 

I applied it several years ago, but while almost nobody interest it.

 

Currently all minimal phase filters is 2G (second generation).

 

Linear phase filters for sample rates D64(2.8 MHz)/128(5.6 MHz) /256(11.2 MHz)/512(22.5 MHz) and 44.1 kHz (after yesterday) replaced to 3G.

 

3G filters technically has less noise/distortions, sometime faster.

 

Technically applying 3G filters to minimal phase filter is possible.

 

I while don’t touched min. phase filters due a few interest.

 

However me seem filters without pre-ringing deserve attention.

 

Me very interesting discussion about this almost unexplored part of audio conversion.

 

Best regards,

Yuri

 

Yuri -

I have not been keeping up with the updates, but I think I have 2G.

 

I just want to understand something, there are no settings for the linear phase filter that Jud is referring to. Correct?

 

The only setting I see is the toggle of the minimal phase filter, on or off. Correct?

 

So, when Jud refers to his experimentation with linear phase filter versus minimal phase filter, he is simply toggling the minimal phase filter on or off. Correct?

 

I plan on downloading the 3G soon, but what I have now is very good using D128 setting. I love the soundstaging on my PMC fact.8 speakers. These are not small changes that only audiophiles can hear.

[audiventory]

Hi Augustine,

 

> I have not been keeping up with the updates, but I think I have 2G.

 

You can reload via last version 4.1.22 via last sent me to you link. If link lost, write me, please. I will re-send it to your e-mail.

 

For version 4.1.22 is actual all as I described above.

 

 

 

> I just want to understand something, there are no settings for the linear phase filter that Jud is referring to. Correct?

 

> The only setting I see is the toggle of the minimal phase filter, on or off. Correct?

 

> So, when Jud refers to his experimentation with linear phase filter versus minimal phase filter, he is simply toggling the minimal phase filter on or off. Correct?

 

 

Yes. Now possibly switch between linear/min. phase filter.

 

While I’m use manual tuned (fixed in released software) filters. Therefore it currently havn't settings.

 

Of course I will try new modes. As example extended band 24...25 kHz. Possibly also intermediate filer, as prefer John Swenson.

 

 

However while all it planned as kit of fixed filters.

 

 

> I plan on downloading the 3G soon, but what I have now is very good using D128 setting. I love the soundstaging on my PMC fact.8 speakers. These are not small changes that only audiophiles can hear.

 

Yes. Need know what hear. Even difference between hi-fi and boombox is not clear for man who listened before lo-fi apparatus only. After getting some experience in hi-fi audio, enough subtle sound differences became clear. I know it by own experience and experience other people. Same things with difference between hi-fi and hi-end.

I consider hi-end as hi-fi that able produce transparent sound not only for mid and high loudness but for low, lowest too.

 

Best regards,

Yuri

[Jud]

I personally prefer intermediate phase filters, you can split the difference between what you heard, they have some pre-ringing, but less than linear phase and less post ringing than minimum phase.

 

John S.

 

Hi John, what speakers are you using, and is there anything else in the system that might interact with phase effects or dispersive filters?

 

Edit: By the way, my preference for linear phase is true of iZotope built into Audirvana+ as well.

[Jud]

Hi Jud,

 

Thank you for testing minimal phase filter.

 

I applied it several years ago, but while almost nobody interest it.

 

Currently all minimal phase filters is 2G (second generation).

 

Linear phase filters for sample rates D64(2.8 MHz)/128(5.6 MHz) /256(11.2 MHz)/512(22.5 MHz) and 44.1 kHz (after yesterday) replaced to 3G.

 

3G filters technically has less noise/distortions, sometime faster.

 

Technically applying 3G filters to minimal phase filter is possible.

 

I while don’t touched min. phase filters due a few interest.

 

However me seem filters without pre-ringing deserve attention.

 

Me very interesting discussion about this almost unexplored part of audio conversion.

 

Best regards,

Yuri

 

Yes, I'd be interested in hearing whether a third generation minimum phase filter would sound more like, or perhaps even preferable to, the linear phase filter.

[Paul R]

I think you have cogently described something that is virtually intangible, and are pretty much correct in your thinking. At least, I agree with you.

[rodrigaj]

Hi Augustine,

 

> I have not been keeping up with the updates, but I think I have 2G.

 

You can reload via last version 4.1.22 via last sent me to you link. If link lost, write me, please. I will re-send it to your e-mail.

 

For version 4.1.22 is actual all as I described above.

 

 

Best regards,

Yuri

 

Yuri -

Thank you. I have the link saved. I'll download and try it.

 

Augustine

[audiventory]

Yes, I'd be interested in hearing whether a third generation minimum phase filter would sound more like, or perhaps even preferable to, the linear phase filter.

 

After replacing to 3G rest linear phase filter, I will work under minimal phase filters.

[Tony Lauck]

Yes, I'd be interested in hearing whether a third generation minimum phase filter would sound more like, or perhaps even preferable to, the linear phase filter.

 

If you can hear differences in filters when upsampling from 192 kHz it's a strong indication that the sampling rate of 192Khz is still marginal, assuming that the filters you compared were not botched.

 

These differences will be greater when upsampling from 44.1 kHz. I found that recordings (that I made from downsampling hi-res) that used minimum phase sounded better when played back using linear phase and, conversely that recordings made using linear phase filters sounded better when played back with minimum phase.

 

I suggest trying a variety of recordings before reaching a firm conclusion as to your preference.

[Jay-dub]

IIRC, all the standard filters of AuI use cutoffs close to 20kHz even when resampling between rates much higher than 44.1. Thus audible differences between the filters are not surprising at all, and are no indication that sampling rates of 192kHz might be marginal. Whether that design constitutes botched filters is a separate question. It is not my preference. I suspect that Jud's system, like many, benefits from content that is band-limited to the 20-kHz audio band, and that this is the main reason he prefers the sound when his recordings are first processed through AuI. Personally, I prefer to stick with a resampling tool that keeps a bandwidth of at least 90% of the Nyquist frequency, and a separate filter that addresses system limitations and personal preference irrespective of the DAC's native rate.

 

When it comes to personal preference on the 20-kHz lowpass, I'm about the same as Jud. For all the recordings I download in high sample rates, I make versions downsampled to 48kHz with a phase-linear filter, and so most of what I listen to on speakers has gone through the the same lowpass filter, which sounds good to me and which I am extremely used to. This avoids potential problems from the break-up frequency of my aluminum dome tweeters. The high-bandwidth files I reserve for headphone listening and for a future system. Phase-linear lowpass filters just sound closer to transparent. They affect the transients slightly, but do not change the tone color. (Barry Diament has written that minimum-phase filters "darken" the tone unacceptably; this is as good a way of putting it as any.)

[Jud]

Thanks Tony and Jay-dub. FYI, I prefer AuI's offline filtering played back with Audirvana+ slightly to inline filtering with Miska's HQPlayer. HQPlayer is at somewhat of a disadvantage because of my 2009 vintage laptop. Also, I'm not using an NAA. I prefer both of these to A+ inline using my own settings for iZotope.

 

That's all on OS X. On Windows I like XXHighEnd, and on Linux of course HQPlayer.

 

Re minimum vs. linear phase, wherever I've compared them - iZotope, HQPlayer, AuI, the alternate filters offered for the Geek Out - I prefer linear phase.

[audiventory]

IIRC, all the standard filters of AuI use cutoffs close to 20kHz even when resampling between rates much higher than 44.1. Thus audible differences between the filters are not surprising at all, and are no indication that sampling rates of 192kHz might be marginal. Whether that design constitutes botched filters is a separate question. It is not my preference. I suspect that Jud's system, like many, benefits from content that is band-limited to the 20-kHz audio band, and that this is the main reason he prefers the sound when his recordings are first processed through AuI. Personally, I prefer to stick with a resampling tool that keeps a bandwidth of at least 90% of the Nyquist frequency, and a separate filter that addresses system limitations and personal preference irrespective of the DAC's native rate.

 

When it comes to personal preference on the 20-kHz lowpass, I'm about the same as Jud. For all the recordings I download in high sample rates, I make versions downsampled to 48kHz with a phase-linear filter, and so most of what I listen to on speakers has gone through the the same lowpass filter, which sounds good to me and which I am extremely used to. This avoids potential problems from the break-up frequency of my aluminum dome tweeters. The high-bandwidth files I reserve for headphone listening and for a future system. Phase-linear lowpass filters just sound closer to transparent. They affect the transients slightly, but do not change the tone color. (Barry Diament has written that minimum-phase filters "darken" the tone unacceptably; this is as good a way of putting it as any.)

 

Hi Jay-dub,

 

As rule nobody don't listen upper 17-20 kHz.

 

However limiting band to 20...22 kHz (usual transient band for AuI) allow us:

 

1. Avoid intermodulation distortions (when harmonics from ultrasound shifted/mirrored to audible range)

 

2. Unload wide band energy, that give for audible frequency range more wide dynamic.

 

Example:

If we have in audible range 1 sine signal with maximal amplitude.

Any increasing of level give us overloading.

 

Next to 1 sine audible sine with maximal amplitude we add 1 ultrasound sine with even low amplitude:

We get distorted sound in audible range.

It's intermaodulations that shifted to audible range.

As example:

 

24 kHz (ulitrasound sine) - 13 kHz (sine in audible range) = 11 kHz (distortion sine)

24 kHz (ulitrasound sine) - 2*9 kHz (sine in audible range) = 6 kHz (distortion sine)

47 kHz (ulitrasound sine) - 3*15 kHz (sine in audible range) = 2 kHz (distortion sine)

 

 

Resume:

Analog part get audible and ultrasound energy as solid energy.

 

If we removed non-audible energy (limit band over 20 kHz), we got reserve for using full potencial of audible range.

 

Other words, in audible range possible playback more intensive parts of music pieces without distortions and simultaneously listen weak parts above noise.

 

Best regards,

Yuri

[Jay-dub]

Thanks, Yuri; that's a rationale.

 

It seems there are three issues at play here:

1. the validity of the theory that human hearing is limited to no more than 17-20 kHz;

2. the amount of additional dynamic range to be had by filtering out ultrasonic content in recorded music files;

3. the appropriateness of this filtering in a general-purpose SRC tool.

 

On point 1, let's just say I have my doubts. So do many other users on this forum. Some believe that even a 40-kHz lowpass causes an audible degradation to a music signal.

 

On the second point, the amount of additional headroom you will get to increase volume without distortion depends on the amount of ultrasonic content on the recording. If you're dealing with a transfer from analogue tape or vinyl, it might be substantial, but the music I've got recorded in 24/96 has very little energy above 20 kHz, and it consists of overtones and other noises from the musical instruments. I doubt you could even get 0.3 dB extra headroom by filtering that out.

 

The third question is one of design philosophy. By putting a 20-22 kHz lowpass in your SRC, you are taking a tool that could operate by completely general mathematical principles and making it less general, out of considerations connected to a psychoacoustic theory and assumptions about the nature of the input signals and the use of the output. This is what I describe as "not my preference." The benefits you ascribe to the filtering could equally well be achieved by the user prefiltering the input.

[audiventory]

Thanks, Yuri; that's a rationale.

 

It seems there are three issues at play here:

1. the validity of the theory that human hearing is limited to no more than 17-20 kHz;

2. the amount of additional dynamic range to be had by filtering out ultrasonic content in recorded music files;

3. the appropriateness of this filtering in a general-purpose SRC tool.

 

On point 1, let's just say I have my doubts. So do many other users on this forum. Some believe that even a 40-kHz lowpass causes an audible degradation to a music signal.

 

As audio engineer I can’t operate terms like believe. Anyway I always keep in mind same things due possibly now I don’t know anything.

However while not rigid proofs of impacting ultrasound to musical perception. May be in future theory will altered.

While we have stable proof that human ear listen up to 20 kHz.

 

 

On the second point, the amount of additional headroom you will get to increase volume without distortion depends on the amount of ultrasonic content on the recording. If you're dealing with a transfer from analogue tape or vinyl, it might be substantial, but the music I've got recorded in 24/96 has very little energy above 20 kHz, and it consists of overtones and other noises from the musical instruments. I doubt you could even get 0.3 dB extra headroom by filtering that out.

 

In DXD signal we have significant ultrasound level of noise http://www.realhd-audio.com/?p=1739

 

Below I try estimate energetic reserve for audible signal.

 

1. We have sample rate 192 kHz.

2. Thus full signal band 96 kHz.

3. Useful signal band 20 kHz.

4. Full energy 1.0

3. Useful energy 20/96=0.208

5. Ultrasound energy 1.0 - 0.208 = 0.792

6. Reserv of energy for useful (audible) signal 10*log(0.792/0.208)=5.8 dB (10 here due it’s power of signal, not voltage).

 

It’s almost equal the headroom (6 dB). I.e. it like (if I don't mistaken in calculations) we can work without headroom and use full bit depth. Or decrease headroom value.

 

It’s very usual for quiet fragments of classical music.

 

 

 

The third question is one of design philosophy. By putting a 20-22 kHz lowpass in your SRC, you are taking a tool that could operate by completely general mathematical principles and making it less general, out of considerations connected to a psychoacoustic theory and assumptions about the nature of the input signals and the use of the output. This is what I describe as "not my preference." The benefits you ascribe to the filtering could equally well be achieved by the user prefiltering the input.

 

AuI is resample for audio applications. It adapted to ears. Resampler for telecommunications usually must support band as half of sample rate.

 

Not enough apply 20-22 kHz filter. It should be professional filter.

 

For offline conversion (not limited by time) filter (desirable) should be near theoretical («ideal») model.

 

AuI use manually and individually tuned filters for each combination input/output sample rate. It allow visual control of full combination of filter’s features: frequency and phase responses, ringing, speed performance.

 

Using filter 20-22 kHz is not my last work.

 

Currently 3rd generation filters releases.

 

Last time I not once get suggestions about apply 25 kHz band.

 

I think do it as option and experiment. However it again will manually tuned filter.

 

Now I apply all filters with 96 kHz at input or output (except already released D64…512).

 

I apply there transient band 20 … 24 kHz. We can compare it with old version by hearing.

 

It will interesting for me.

[Tony Lauck]

As audio engineer I can’t operate terms like believe. Anyway I always keep in mind same things due possibly now I don’t know anything.

However while not rigid proofs of impacting ultrasound to musical perception. May be in future theory will altered.

While we have stable proof that human ear listen up to 20 kHz.

 

 

 

 

In DXD signal we have significant ultrasound level of noise What About DXD? Surprise! | Real HD-Audio

 

Below I try estimate energetic reserve for audible signal.

 

1. We have sample rate 192 kHz.

2. Thus full signal band 96 kHz.

3. Useful signal band 20 kHz.

4. Full energy 1.0

3. Useful energy 20/96=0.208

5. Ultrasound energy 1.0 - 0.208 = 0.792

6. Reserv of energy for useful (audible) signal 10*log(0.792/0.208)=5.8 dB (10 here due it’s power of signal, not voltage).

 

It’s almost equal the headroom (6 dB). I.e. it like (if I don't mistaken in calculations) we can work without headroom and use full bit depth. Or decrease headroom value.

 

It’s very usual for quiet fragments of classical music.

 

 

 

 

 

AuI is resample for audio applications. It adapted to ears. Resampler for telecommunications usually must support band as half of sample rate.

 

Not enough apply 20-22 kHz filter. It should be professional filter.

 

For offline conversion (not limited by time) filter (desirable) should be near theoretical («ideal») model.

 

AuI use manually and individually tuned filters for each combination input/output sample rate. It allow visual control of full combination of filter’s features: frequency and phase responses, ringing, speed performance.

 

Using filter 20-22 kHz is not my last work.

 

Currently 3rd generation filters releases.

 

Last time I not once get suggestions about apply 25 kHz band.

 

I think do it as option and experiment. However it again will manually tuned filter.

 

Now I apply all filters with 96 kHz at input or output (except already released D64…512).

 

I apply there transient band 20 … 24 kHz. We can compare it with old version by hearing.

 

It will interesting for me.

 

Filtering out ultrasonic noise may reduce the wide-band noise level, but not a weighted noise level related to hearing thresholds.

 

Headroom is determined by peak signal amplitudes. Intuitively, if one filters out certain frequencies one might think that headroom would improve, but this intuition would be wrong. Filtering out frequencies can actually increase peak signal amplitudes. For example, if one has a square wave at peak amplitude of V, and one filters out all of the harmonics of the square wave, one will get a sine wave that has a peak amplitude of (4/pi)*V, about 2 dB greater than the original signal.

 

If a 192 kHz recording has energy in the range from 24 to 80 kHz this is part of the original recording and a high quality DAC should output the recording the way it was made. Filtering this out amounts to distorting the original recording, e.g. removing some of the "bite" on brass instruments. This may be a good idea in a DAC aimed at the mid-fi market where it is presumably going to be used with mid-fi amplifiers that are unable to handle input signals with significant ultrasonic energy, but this is would completely defeat the purpose of going to the 192 kHz sampling rate. If everything at and above 24 kHz is going to be completely filtered out in playback, then one can get the exact same results by down sampling the 192 kHz to 48 kHz.

 

Some recordings may actually sound better when high frequencies are filtered out. (For example, the microphone preamplifiers might have picked up some ultrasonic noise from video circuitry.) If this is the case, then the problem should have been caught in the production process, ideally by the tracking engineer, but if necessary caught and fixed by the mastering engineer. If the problem has to be fixed after the musicians are gone, then it will be a judgment call on the part of the mastering engineer, since one will be trading off different sound qualities. This is not the kind of judgment that a DAC designer can possibly make.

[Jud]

Re ultrasonics, are we talking about conversion to PCM or DSD, or a PCM->PCM step on the way to DSD...?

[audiventory]

Filtering out ultrasonic noise may reduce the wide-band noise level, but not a weighted noise level related to hearing thresholds.

 

Due intermodulation distortion in analog part ultrasonic noise mirrored/shifted to audible range. More wide range - more noise energy able for shifting.

 

 

Headroom is determined by peak signal amplitudes. Intuitively, if one filters out certain frequencies one might think that headroom would improve, but this intuition would be wrong. Filtering out frequencies can actually increase peak signal amplitudes. For example, if one has a square wave at peak amplitude of V, and one filters out all of the harmonics of the square wave, one will get a sine wave that has a peak amplitude of (4/pi)*V, about 2 dB greater than the original signal.

 

Square wave has more energy in spectrum than sine. So square achieve overloading with lesser maximal amplitude than sine.

 

If a 192 kHz recording has energy in the range from 24 to 80 kHz this is part of the original recording and a high quality DAC should output the recording the way it was made. Filtering this out amounts to distorting the original recording, e.g. removing some of the "bite" on brass instruments. This may be a good idea in a DAC aimed at the mid-fi market where it is presumably going to be used with mid-fi amplifiers that are unable to handle input signals with significant ultrasonic energy, but this is would completely defeat the purpose of going to the 192 kHz sampling rate. If everything at and above 24 kHz is going to be completely filtered out in playback, then one can get the exact same results by down sampling the 192 kHz to 48 kHz.

 

Unfortunately, not robust proof what we listened all what upper 20 kHz.

 

High sample rates used not for playback inaudible range. But for applying more simple analog filter in DAC, even hi-end.

 

After downsampling we must use more complicated filters for achieve result like to using high sample rate.

 

 

Some recordings may actually sound better when high frequencies are filtered out. (For example, the microphone preamplifiers might have picked up some ultrasonic noise from video circuitry.) If this is the case, then the problem should have been caught in the production process, ideally by the tracking engineer, but if necessary caught and fixed by the mastering engineer. If the problem has to be fixed after the musicians are gone, then it will be a judgment call on the part of the mastering engineer, since one will be trading off different sound qualities. This is not the kind of judgment that a DAC designer can possibly make.

 

Some AuI users convert certain (I suppose, it's DXD and derived form it) high-resolution files without resampling. Simple for removing of noise.

 

Simple example video about playback such high resolution file

[audiventory]

Archimago suggested like the thread's topic Archimago's Musings: INTERNET BLIND TEST: Linear vs. Minimum Phase Upsampling Filters