Apodizing filter

[noahyd]

im sure if i follow your methadology i end up at 50/50 again, tho not because its "the only truth", but that just comes from my book

I feel like you'd be a good participant in drug trials.

Ghoost: "Doctor, I've had this pain in my shoulder from the medication."

Doctor: "We'll keep you under observation."

Ghoost: "But doctor... it hurts."

Doctor: "You're on the placebo."

Ghoost: "That's not the only truth in my book."

Doctor: "Oka-"

Ghoost: "By the way, the pain is gone. That was awful; dial back on my dosage."

 

[noahyd]

Yeah

Fun fact: You joined this forum when I was 1 year old.

That's quite cool.

Anyway, I shall not derail the thread...

 

[danadam]

Fun fact:

Maybe to you

 

[Ghoostknight]

Isn't he referring to your MP3 320kbps vs. Flac test?

it even makes less sense then, well, atleast i suspect audacity to keep the same volume...

 

[noahyd]

it even makes less sense then, well, atleast i suspect audacity to keep the same volume...

It makes sense. Higher volume is perceived as more quality. Small volume changes can account for the test results, as he said, if the MP3 files' volume changed during converting, and the other files did not.

But yes, if you did not make changes to it in Audacity, volume should be the same. But it still uses FFMPEG and LAME encoder, I think, and maybe it has some audible effects such as volume change that I don't know of.

That said, I don't comment on your test results; just clarifying my understanding on the volume thing.

 

[gregorio]

sure inaudible frequencys but the volume changes, alright m8

All the different encoders have to change the volume! The process of encoding requires up/oversampling and oversampling will commonly cause clipping distortion (due to inter-sample peaks). That’s not an issue for DACs because all oversampling DACs provide headroom specifically to address this issue of ISPs, usually at least 3dB headroom and sometimes as much as 6dB but of course software upsampling has to lower the volume to provide this headroom. Typically the software will then compensate and raise the level again at the end of the encoding process, however there is often still a slight difference (a few tenths of a dB). Now that’s not enough for anyone to notice with just casual listening but it can be enough to affect the results of ABX tests in some cases and therefore needs to be compensated for manually. Being ignorant of this fact when arguing about it in a science discussion forum is bad enough but your sarcastic tone just makes you look even more foolish!

im sure if i follow your methadology i end up at 50/50 again

Me too, I’m pretty sure that if you eliminate the biases, along with the variables you’re not testing for, then you’ll probably end up at 50/50 on average. Although it is still possible to get a 70% result just by pure chance.

tho not because its "the only truth", but that just comes from my book

Sure, but as you state that “just comes from your book” which you’ve demonstrated repeatedly is a book where you make up BS, defend it with fallacies and ignore all the science/facts that proves it’s BS. However, to science and any rational mind then it would be at least a “truth” if the test does indeed eliminate those variables/bias effects that would otherwise produce incorrect results. It’s not a difficult concept to grasp or accept, even a relative idiot should understand that not eliminating other potential causes will not provide any sort of “truth”! Of course though, this excludes someone who has made-up BS that’s contrary to the facts and their ego won’t let them see/accept the facts/truth!

But yes, if you did not make changes to it in Audacity, volume should be the same.

There should be a volume change, even ignoring the ISP (inter-sample peak) issue mentioned above. A lossy codec is removing frequency content, so obviously that will affect the volume level. However, freqs are removed according to auditory masking and other psychoacoustic principles and should therefore be inaudible, even though the level will be slightly/somewhat different. It used to be relatively easy to differentiate MP3 320 from a lossless original but that was around 25 years ago when the psychoacoustic models employed were still relatively poor/basic but they were improved over time and by about 15 years or so ago it was impossible to tell the difference at normal listening levels. There are still a handful of tracks that can be differentiated at 320kbps (due to a pre-echo issue if I recall correctly) but they are rare/uncommon tracks and “Smells Like Teen Spirit” is not one of them.

But it still uses FFMPEG and LAME encoder, I think, and maybe it has some audible effects such as volume change that I don't know of.

AFAIK, LAME is still bundled with Audacity and if I remember correctly it was the first MP3 encoder to start using better psychoacoustic models but we don’t know if Ghost used a modern version of the LAME encoder or what settings he used if so. If you’re interested in this stuff, head over to the “Hydrogen Audio” forum, that’s where the LAME developer hangs out and where a lot of the testing occurs, or at least it was when I was still interested in improved MP3 encoders.

G

 

[Ghoostknight]

It makes sense. Higher volume is perceived as more quality. Small volume changes can account for the test results, as he said, if the MP3 files' volume changed during converting, and the other files did not.

But yes, if you did not make changes to it in Audacity, volume should be the same. But it still uses FFMPEG and LAME encoder, I think, and maybe it has some audible effects such as volume change that I don't know of.

That said, I don't comment on your test results; just clarifying my understanding on the volume thing.

i went ahead and exported the original .flac trough audacity two times
1. as MP3 (Insane/320kbit)
2. as WAV

just so any potential "processing" by audacity is applied to both files im testing, not sure what i have could done else here

tho its probably easy to check peak levels in audacity afterwards, i still have both files

What will happen if i run a nulltest between a MP3 and WAV?, atleast from what i hear i would suspect deep bass being different and high frequencys above 15khz or so


@gregorio thanks for clarifying on this one, i wasnt aware of this, let me check peak levels on the original (EDIT: it was -0,26db) , so your suggesting on these kind of tests is having maximum -3db peaks?

its Audacity 3.4.2 on ubuntu, tho i have the ubuntu package installed and lame isnt installed, not sure about that
EDIT:
So, the lame package is probably packaged in some other package (ffmpeg or libav isnt either installed, see the pic below from the audacity wiki)

but since audacity shows the lame presets like "insane" im pretty sure it uses lame in the background

 

[gregorio]

i went ahead and exported the original .flac trough audacity two times
1. as MP3 (Insane/320kbit)
2. as WAV
just so any potential "processing" by audacity is applied to both files im testing, not sure what i have could done else here

You would obviously get different results. Converting flac to wav is just a relatively simple process of converting substituted mathematical symbols back into the exact same data values you started with, there is no audio processing occurring, therefore no headroom or any other audio changes. This is not the case with Lossy codecs, which employ a lot of audio processing. However, it should ALL be inaudible at high bitrates.

What will happen if i run a nulltest between a MP3 and WAV?, atleast from what i hear i would suspect deep bass being different and high frequencys above 15khz or so

This is the one example where a null test provides no/little useful information. Because there are significant spectral differences between a MP3 and WAV due to the fact that the MP3 encoder will be removing freqs that are masked and as masked frequencies can occur throughout the audible spectrum we will get a null test difference file with differences throughout most/all of the spectrum and that’s in addition to the slight volume differences we’re also likely to encounter with the encoder. However, the bass will typically be the least affected/different because the number of freqs masked increases the higher in the spectrum you go. The encoder will also probably be removing freqs above about 16kHz (which are inaudible to adults) but not necessarily, it depends on the exact settings of the encoder.

@gregorio thanks for clarifying on this one, i wasnt aware of this, let me check peak levels on the original (EDIT: it was -0,26db) , so your suggesting on these kind of tests is having maximum -3db peaks?

No, the level will be lowered internally by modern MP3 encoder software (and then raised again at the end of the process), you don’t need to do anything and flac -> WAV does not incur any volume changes, it’s identical. The encoded MP3 may end up with exactly the same peak levels (or there maybe a slight difference in peak levels) but may have a slightly different RMS or “Loudness” level, there’s no set amount of how much volume or peak level difference there will (or will not) be, it varies according to encoder’s analysis of the content and of course the content is different for each individual track encoded. Check for a difference in peak and RMS levels, and with a loudness meter and compensate one or other of the files until the loudness is the same.

G

 

[Ghoostknight]

This is the one example where a null test provides no/little useful information. Because there are significant spectral differences between a MP3 and WAV due to the fact that the MP3 encoder will be removing freqs that are masked and as masked frequencies can occur throughout the audible spectrum we will get a null test difference file with differences throughout most/all of the spectrum and that’s in addition to the slight volume differences we’re also likely to encounter with the encoder. However, the bass will typically be the least affected/different because the number of freqs masked increases the higher in the spectrum you go. The encoder will also probably be removing freqs above about 16kHz (which are inaudible to adults) but not necessarily, it depends on the exact settings of the encoder.

yea thought so, tho since i hear a difference (atleast im pretty sure) it would be still somewhat interesting to see a spectrogram of the difference

No, the level will be lowered internally by modern MP3 encoder software (and then raised again at the end of the process), you don’t need to do anything and flac -> WAV does not incur any volume changes, it’s identical. The encoded MP3 may end up with exactly the same peak levels (or there maybe a slight difference in peak levels) but may have a slightly different RMS or “Loudness” level, there’s no set amount of how much volume or peak level difference there will (or will not) be, it varies according to encoder’s analysis of the content and of course the content is different for each individual track encoded. Check for a difference in peak and RMS levels, and with a loudness meter and compensate one or other of the files until the loudness is the same.

ah alright, i just thought i could manually reduce peak levels to avoid the reamplifying, probably not making a difference but still, may be less processing overall, but sounds like i should just let mp3/lame do its thing


MP3:
Peak:
-0.273db
RMS:
Left: -16.2022 dB
Right: -15.839 dB
Stereo: -16.0168 dB

WAV:
Peak:
-0.337db
RMS:
Left: -16.2036 dB
Right: -15.8399 dB
Stereo: -16.0179 dB
(i checked the original flac file too, its indeed identical)

its pretty close, everything is under around 0.05db difference

did mp3 increase peak levels to compensate RMS values ? (because stuff gets filtered out, reducing RMS in theory)

i guess we can say its "inaudible" tho it might "blur" the result just a tiny bit...

 

[gregorio]

ah alright, i just thought i could manually reduce peak levels to avoid the reamplifying, probably not making a difference but still, may be less processing overall

AFAIK, that’s an automatic process within the encoder, so lowering the volume before encoding wouldn’t make any difference. The MP3 encoder just processes a “window” (or a few combined “windows”) of audio data at a time, it doesn’t analyse the whole track/file first AFAIK, so it wouldn’t know the peak level is say -3dB until it got to the end of the track.

its pretty close, everything is under around -0.05db difference

The peak level is different by 0.064dB according to your figures, so obviously there is a change in levels, although it’s well below audibility in this instance.

did mp3 increase peak levels to compensate RMS values ? (because stuff gets filtered out, reducing RMS in theory)

I assume so, although it obviously wouldn’t be able to do that in cases where the amount of compensation required for the RMS level would put the peak level at or above 0dBFS and then there would likely be several tenths of a dB difference in RMS levels. Older encoders did in fact try to do that and therefore could sometimes be differentiated due to the clipping that would result. That was always one of the benefits of the AAC encoder, which was less prone to clipping but again, that’s only a potential issue with older MP3 encoders.

i guess we can say its a "inaudible" tho it might "blur" the result just a tiny bit...

Yep, 0.064dB is definitely inaudible, even over a long duration, let alone just for a few short peaks. I’m not sure what you mean by “blur”, that’s a poorly defined marketing term. If you mean pre-echo, then yes it can occur to audible levels with MP3 but there’s only one or two tracks these days where this is the case. I seem to remember there was a traditional Spanish music track with single castanet sections where it could be reliably detected.

G

 

[Ghoostknight]

I’m not sure what you mean by “blur”, that’s a poorly defined marketing term. If you mean pre-echo, then yes it can occur to audible levels with MP3 but there’s only one or two tracks these days where this is the case. I seem to remember there was a traditional Spanish music track with single castanet sections where it could be reliably detected.

more often than not you hear a difference but cant put your finger on it what its, thats for me "blur" but i generally wouldnt worry about -0.064db, but since we are talking MP3 vs Lossless, where differences are small to begin with i wouldnt necessarly exclude the volume difference completely here

also what shouldnt be forgotten, i surely could compensate RMS values afterwards, but who is doing this in "real world use" ? i kinda think the mp3 codec should be taken as what its and does

I assume so, although it obviously wouldn’t be able to do that in cases where the amount of compensation required for the RMS level would put the peak level at or above 0dBFS and then there would likely be several tenths of a dB difference in RMS levels. Older encoders did in fact try to do that and therefore could sometimes be differentiated due to the clipping that would result. That was always one of the benefits of the AAC encoder, which was less prone to clipping but again, that’s only a potential issue with older MP3 encoders.

interesting, yes makes sense

i imagine the "windowing" makes it impossible to use the RMS compensation agressively or it would be kind of obvious (like an "pumping" compressor, sort of)

Yep, 0.064dB is definitely inaudible, even over a long duration, let alone just for a few short peaks. I’m not sure what you mean by “blur”, that’s a poorly defined marketing term. If you mean pre-echo, then yes it can occur to audible levels with MP3 but there’s only one or two tracks these days where this is the case. I seem to remember there was a traditional Spanish music track with single castanet sections where it could be reliably detected.

pre-echo == pre-ringing, right? i guess this comes from the above 16khz (lowpass?), where we would end up again at my Lowpass blind test or what this thread is about, reconstruction filters that behave similar to a lowpass

 

[gregorio]

more often than not you hear a difference but cant put your finger on it what its, thats for me "blur" but i generally wouldnt worry about -0.064db, but since we are talking MP3 vs Lossless, where differences are small to begin with i wouldnt necessarly exclude the volume difference completely here

The human threshold is at best ~0.2dB and even then, only under optimal conditions (not short peaks), this is why DBT/ABX requires volume matching to 0.1dB. 0.064dB is significantly below even the conservative 0.1dB requirement of ABX, so yes, it should necessarily be excluded here!

also what shouldnt be forgotten, i surely could compensate RMS values afterwards, but who is doing this in "real world use" ?

Almost everyone. Don’t you ever change the volume for different tracks/albums to get the best volume? Sure, no consumer is going to make adjustments for just a few tenths of a dB which they won’t notice but again, it is sometimes required when ABX testing because a few tenths of a dB can be audible/detectable, especially when fast switching short sections, and therefore should be checked and compensated for to avoid incorrect results.

pre-echo == pre-ringing, right? i guess this comes from the above 16khz (lowpass?), where we would end up again at my Lowpass blind test or what this thread is about, reconstruction filters that behave similar to a lowpass

No, pre-echo has nothing to do with pre-ringing. Pre-ringing is a consequence of a linear phase filter in response to an impulse and results in a brief oscillation before the impulse at/near the Nyquist Frequency. Pre-echo on the other hand is a consequence of the overlapping processing of the transform (MDCT) based windows used by MP3 and some other lossy encoders and results in the sound (echo) being heard before the transient that caused it (and typically the transient will therefore be masked or partially masked). It is not an oscillation (ringing), does not occur only near the Nyquist frequency (it contains roughly the same freqs as the transient) and has nothing to do with anti-imaging/reconstruction filters! With a tiny number of exceptions (<5), this potential problem has effectively been eliminated over the years as pre-echo mitigation methods have been developed and implemented. Some encoders (Vorbis for example) include parameters in their advanced options that combat pre-echo if you ever encounter it.

G

 

[Ghoostknight]

Almost everyone. Don’t you ever change the volume for different tracks/albums to get the best volume? Sure, no consumer is going to make adjustments for just a few tenths of a dB which they won’t notice but again, it is sometimes required when ABX testing because a few tenths of a dB can be audible/detectable, especially when fast switching short sections, and therefore should be checked and compensated for to avoid incorrect results.

im not sure how much the RMS value can change in extreme cases but imo this shows a "flaw of the codec" and shouldnt be "compensated" solely for a DBT

secondly, im unsure how the discrapency between rms values exactly get created but imagine following:
1. mp3 removes high frequency (theoretically) inaudible content and reduces therefore RMS
2. if you raise now the overall volume to match the RMS values are you not essentially boosting lower frequencys in comparison actually creating a bigger difference?

 

[bigshot]

Fun fact: You joined this forum when I was 1 year old.

I just noticed I had my 20th anniversary on Head Fi a couple of weeks ago.

Happy 21st birthday to you! Cheers!

 

[noahyd]

I just noticed I had my 20th anniversary on Head Fi a couple of weeks ago.

Happy 21st birthday to you! Cheers!

Thank you! Happy Headfiversery!