Sunday 26 April 2015

INTERNET BLIND TEST: Linear vs. Minimum Phase Upsampling Filters


I. Introduction

Ladies and gentlemen, audiophiles of all ages; you have probably seen pictures like the ones above many times in the magazines, posted online, etc. You know that many DACs these days feature the ability to adjust the way the antialiasing reconstruction filters are "tuned". I remember reading about and becoming fascinated by the different filters back in 2006 with Keith Howard's article on Stereophile. But it wasn't until around 2009 that I noticed Meridian started using minimum phase settings in their equipment like the 808.2 Signature Reference CD player. (I don't know if many other consumer audio devices used a form of the minimum phase filter before 2009; perhaps not as well advertised as Meridian?)


As you can see by the images above, when we send a 16/44 "impulse" through these reconstruction filters, we create different forms of high-frequency ringing. Linear phase filters do not result in phase distortion. However, the "issue" we are to note looking at the typical linear phase impulse response is the symmetrical ringing as seen above. It has been suggested that the pre-ringing ("pre-echo") is damaging to the sound and "unnatural". As a result, minimum phase filter algorithms became staples of companies like the aforementioned Meridian and later on Ayre with their well known whitepaper on their filter choice (I see this is also used in the recent PonoPlayer). Realize however that in removing the pre-echo, minimum phase filters will result in high frequency phase shift compared to the linear phase filter as shown here (from the Infinite Wave website looking at SoX 14.4 VHQ Minimum Phase vs. Linear phase):
Phase shift with minimum phase filter.
For completeness: Notice how sharp the SoX filters are which results in the ringing above. Both linear and minimum phase settings show a sharp "brick wall" frequency response like this.

As discussed on this blog and elsewhere, many have criticized the 44.1kHz sampling rate chosen as the CD standard decades ago. The Nyquist frequency of 22.05kHz is said to be too close to the audible frequency typically noted as going up to ~20kHz. Could this "ringing" at 22.05kHz cause audible effects? Can the linear phase "pre-echo" lead to a less realistic sound? Can the minimum phase filter setting sound "better" because we've removed the pre-echo (the post-echo ringing supposedly masked and less likely to cause audible effects - post-masking) despite the phase shift?

Remember that the technical discussion of digital filters can get complicated, often with no actual data around audibility to support positions. Over the years, special player software such has HQPlayer have been released with upsampling algorithm tweaking as a major feature. However, I hope that in using the open-source SoX algorithm with basic parameters that are not specially selected to reduce the "ringing", this test can provide some insight into audibility in the "real-world" with your help.

Here's your chance to listen/test for yourself and let me know what you hear!

II. Test Procedure

As usual, I have put the test files to download off my FTP server and thanks again to Ingemar at Privatebits, the file can be downloaded there as well. Warning: File size of 130MB, "Linear-vs-Minimum-Phase.zip".

Download off Archimago's FTP (digitalfilterstest.dyndns.org) - Login: "filter", Password: "test"

Download off Privatebits

As you see in the ZIP package, there are six FLAC files; each ~1 minute in length named:

A big thank you to Juergen for providing the first two excellent sounding recordings (Mandolin and GrandPiano) as 24/44 WAV files. As Juergen described when he provided the samples, the recordings were done in pure stereo using 2 cardioid condenser microphones, equivalence stereophony technique. No dynamic compression, no limiting, no reverb or other artificial processing. The music was originally recorded at 24/88 and downsampled with Reaper. I'll let Juergen describe the recordings specifically:

Mandolin concerto: "recorded live in a church. So lots of room information and a bit longer decay in the bass section. But fine and nice details."

GrandPiano: "And here is a photo showing the microphones, where they are placed, during the recording of the Steinway Grand D Piano. This is the only Grand D in Europe that is not painted with piano lacquer, but only with some thin matte black, in order to have a more raw sound. This is the Piano where Oscar Peterson played on, on some of his recordings. This piano has a more 'wooden' sound and not so 'polished' as the Grand D in concert houses."



The third sample is the title track from Mighty Sam McClain's "Give It Up To Love" (AudioQuest, 1992); a very well recorded blues album from an artist who I believe is well known among audiophile circles. This is certainly one of those "demo CD's" worth showing off a hi-fidelity system with. Original source is 16/44 CD rip.

As usual, I use these samples under the principle of "fair use" for the purpose of research and commentary.

I took ~1 minute of these three 24/44 or 16/44 recordings and using SoX, upsampled them to 24/176.4 with either the linear or minimum phase upsampling algorithm, re-naming one file A and another B. Note that A and B are not the same for the three samples, so if I thought "Mandolin A" was the linear phase filter, this does not necessarily mean "GrandPiano A" was also the linear phase output.

Here's the DR measurement result of the test samples:

This confirms that the average amplitude is not significantly different between samples A and B. Slight differences in peak amplitudes (please do not think that this correlates with the type of filtering used!). As you can see, each track maintains excellent dynamic range with DR 12-14.

Your job:
1. Make sure you have the proper pre-requisites:
     - High resolution DAC that can play 24-bit, 176.4kHz sampling rate. (NOTE: Most 192kHz-capable DACs should have no problem.)
     - Make sure your player / DAC is not resampling the audio (eg. don't play on a Squeezebox system that has a maximum of 96kHz for example).
     - Turn off all DSP processing like room-correction and EQ for the computer and if you're using a receiver / processor make sure it's "direct" audio, no Audyssey processing and the like.
     - If you're using playback software like JPlay where you might not be able to verify that it's not adding any processing, I recommend just playing this with foobar and bit-perfect ASIO or 24-bit WASAPI to your DAC.
     - Feel free to use tools like the newest ABX Comparator in foobar. Just make sure again the music is not being resampled and being sent out bit-perfect to the DAC.
     - Recommended to listen with speakers to recreate the soundstage. However, headphones may provide better temporal resolution with no room distortion- your pick, just let me know which!

2. Have a good listen to the A and B samples. Consider which of the 2 samples you liked better - that is, sounding more "real", "lifelike" considering these are natural recordings in "live" space. Which did you think had better sounding transients (better temporal resolution)? Did you hear or suspect you heard any artificial "pre-echo" that would suggest the linear phase filter? Did you hear any high-frequency distortion?

3. Fill out my survey and tell me what you heard. As usual, the data will be collected anonymously but cookies have been turned on to limit submissions from any one machine. Note that in previous blind tests, I have had people contact me asking me what they entered because they forgot. In such cases, it would of course be easier if one could leave some kind of easy-to-find identifier or just make sure you include detail I can easily search for later.

The survey can be found here:

Data collected will include:
- Demographics: gender, age range, which continent you're from
- For each piece of music, which one sounded better / more natural / more realistic? There will also be a choice for "sounds the same" if you don't hear any difference.
- Tell me how confident you are about your choice.
- Describe your audio system - seriously folks, this does NOT need to be some $$$$ system! (Checkbox for headphones vs. speakers for the main system used to evaluate.)
- Subjective impression for what you heard. You might also want to comment if you have special listening training here.
- Identification of whether you're a musician / audio engineer / audiophile writer for cohort analysis if possible.

For the sake of consistency and obtaining a good data set, I will insist that most of the questions must be answered. Thanks!

III. You have 2 months.

Hopefully there will be no rush since the samples are only 1 minute each. The survey will close on JUNE 25, 2015. Make sure to get all your responses in by that time!

Remember, this is a listening test! Please refrain from using an audio editor and your eyes :-). Also remember there is no right or wrong; I just want to get an idea if there is a significant preference out there or if you believe you hear a difference. Also, please do not bias others in discussions... I think it's cool to talk about what you heard or did not hear so perhaps other can learn to perceive the difference but let's not specifically talk about file A or B, and even if you know which is which, let's be courteous and keep this as a "blind" test for others until all data collection is complete. I will of course reveal all when I summarize the results in June or early July.

Feel free to advertise this test among friends / family / enemies. Also, feel free to put this up on the various audiophile websites, Facebook, wherever you know high-resolution audiophiles may reside (remember, these are high-resolution 24/176 test samples). The more testers the better to improve statistical power.

Much appreciated!

----------

I've been getting into some newer music lately. A reader switched me on to Butch Walker's stuff recently so I'm working through his discography. Good "new" rock! Enjoying his 2011 album The Spade. I just wish the recording quality was more natural and less dynamically compressed.

With the weather getting better here in Vancouver, it's time to also get outdoors with the family and enjoy other hobbies like photography; perfect time to focus on other things while data is getting collected! Have fun and let me know in the comments if you run into troubles or have other questions.

Enjoy the music...

UPDATE:
Folks, the test has concluded as planned on June 25. Please find the results starting at Part I: RESULTS; the first of 3 parts with analysis and subjective comments received from the anonymous respondents in the 2nd and 3rd parts.

11 comments:

  1. Last time I watch interest to minimal phase filters. Several years (after applying it in my software) almost nobody ask me about it. I always listened that linear phase filters are more preferable.

    Me very interesting results of the test.

    ReplyDelete
    Replies
    1. Exactly why I'm curious about this!

      Audiophile magazines speak about how listeners can hear the difference so I figure it was time to have a look!

      Delete
  2. Vancouver's one of my favorite cities on Earth, so I'm jealous.

    Probably there's something I'm not getting, but there are at least two reasons I can think of why your test won't test what you think it'll test. :)

    1. In all but a very few DACs, any input below 352.8kHz resolution will be internally upsampled to that resolution or to 384kHz before being sent through a delta-sigma modulator. The result may be a combination of the filtering used to obtain the input file with the DAC's internal filtering, or it may be, particularly in the case of the internal filter being an apodizing filter, that the internal filter characteristics will be heard and not the filtering used to produce the input file.

    In order to be sure what the testers were hearing were the filtering characteristics of the input files, one of two things ought to be true about the input files: (a) They are at 352.8 or 384kHz resolution, in which case people with DACs that accept such resolutions will be able to test; or (b) They are DSD files upsampled from lower resolution PCM, using linear or minimum phase settings, in which case people with DACs that accept DSD input will be able to test.

    2. As noted here - http://www.computeraudiophile.com/f8-general-forum/i-have-theory-about-why-i-linear-phase-filters-24135/#post417480 - my speaker crossovers are designed to preserve phase, and the drivers are "time aligned," which may be responsible for my ability to distinguish blind between minimum and linear phase filters, though not necessarily due to pre-ringing. Minimum phase filters are dispersive and have group delay, which some audio designers (e.g., Keith O. Johnson) feel is audible.

    If dispersion/group delay is at least partly responsible for audible distinctions between minimum and linear phase filters, then the crossover and driver designs of listeners' speakers will play a role in the test results. These might also confound attempts to listen for ringing effects. Is any time domain behavior that is audible due to crossover filter design (in which case it would manifest for both minimum and linear phase inputs), or is it due to the filtering of the input files (or internal filtering of the DAC)? Is transient "smearing" due to drivers not being time aligned, crossover design, ringing, or dispersion/group delay, and if the latter, is it filtering of the input files, or internal DAC filtering?

    Just a whole bunch of confounding factors with these test files as I see it, but again, maybe I'm missing something, and if so I will welcome your explanation.

    ReplyDelete
    Replies
    1. Thanks for the note Jud. Yeah, Vancouver is lovely this time of the year as we head towards warmer days! Hope things are nice where you are (CA?).

      1. Yes, there will be upsampling to 352+ in most DACs. However, SoX has done most of the "heavy lifting" from 44 --> 176kHz already so whatever the DAC is doing internally I suspect will only be a small contribution to the overall effect.

      2. Please have a listen and let me know what you think in the survey! I obviously want to capture folks like yourself with speakers like the Vandersteen to get a sense of how many out there can detect differences... More than this, I think it's important to get a sense of preference (ie. do most people actually prefer the "sound" of minimal or linear?). Also, I think it's important to look at the consistency of respondents... Do they actually have the ability to consistently choose the *same* linear or minimum phase setting across the 3 samples?

      Like you say, many potential variables! I hope that by looking at linear vs. minimum phase steep filter settings like this where the (pre-)ringing is quite pronounced, we can at least get at one of these variables. Obviously if there's no indication at all that folks have a preference, then we really should be talking about these other factors and consider that perhaps digital filtering isn't a big deal and would not be a significant variable in choosing which DACs/settings we prefer...

      Delete
    2. You're a photography hobbyist too?

      I don't have a good enough DAC to do this. I need to upgrade.

      Delete
    3. Yes Ryan, been fooling around with digital photography since 2004... First DSLR was the Nikon D70. These days using either the Nikon D800 or Sony A6000.

      Have a pretty good Epson 13" printer in the basement as well for larger sized printing.

      Not as active on photo forums over the years... Certainly not the kind of contentious discussions as these with audio!

      Delete
  3. Thanks for this test. I can't participate because I need to use FIR convolution filters to listen to my system. I do prefer linear phase filters though for digital crossovers and room correction.

    Vancouver is also my favorite city. :-)

    ReplyDelete
    Replies
    1. Hi Michael!

      Vancouver is nice so long as one can handle the rainy dark winter days - much like Seattle in this respect!

      Yup, wouldn't want to turn on FIR and convolution filters when testing.

      Delete
  4. This comment has been removed by the author.

    ReplyDelete
  5. Pleased to see a test like this, good job.

    Been fascinated with digital filters ever since learning, and discovering by ear digital filters are a compromise. My personal preference is for minimum phase apodizing, whether it's the pre-ringing elimination and/or coloration due to phase distortion, the extra bite and transient attack in the mid-range makes for an enjoyable listen.

    Personally happy that some DACs give users that can hear it a choice of filters for taste, music content, audio equipment.

    Have taken and submitted the test, found it easy to hear the sound characteristics exhibited by the filters used.
    Look forward to the results.

    ReplyDelete
    Replies
    1. Thanks for the submission BArnold!

      I'm looking forward to the final results as well. Certainly objectively the effects are clear to "see" and the filters I'm testing here should give folks as good a chance as any to hear a difference through their system!

      Delete