[Dieucocto]

Yes it cuts some bass response too. The whole point though is that it should be inaudible at the higher bitrates like 256k or 320k. And it should be even harder to spot with more advanced encoders like AAC or Opus.

Ok thank you ! Yeah I know it's supposed to be inaudible, it's really more about how the compression is done that I'm curious. The thing is, if its cuts some bass and mids too, why can't I see it when I compared two spectrums of an 320 Kbits/s and a FLAC ? Like I said, before 16 Khz i see no difference.

 

[castleofargh]

Hi there !

I have a question for you, the compression experts out there !

Here’s the thing : I know, for exemple, that when we compress a flac file into a 320 kbits/s MP3 (with LAME), it cuts everything that’s after the 20 Khz frequence (and sometimes a bit of stuff between the 16 Khz and 20 Khz too, depending on the file). But I was wondering : does it really mean that everything that’s being cut from the MP3 320 kbits/s are the upper treble ? If I’m being logic, yes, but sometimes I’m wondering if it’s that simple, because If that so, it means that even a 128 kbits/MP3 are basically lossless in terms of bass.

To make it short : is the cut of the MP3 compression really that clean (and in that case it really only concerns the trebles), or does it also compress in a way the bass and the mids (even if there’s not really, of course, a clear cut) ? When I compared the spectrum between a 320 kbits/s and a FLAC, I don’t see any difference under the 16 Khz, but I have doubts...
Oh, and I don’t talk about « what can be heard »... I’m really talking about the data itself.

Thank you !

the low bitrate mp3 ends up getting rid of the upper frequencies above 16khz, but not at 320kbps. it can be confusing because the format keeps the same name and we only move a slider to set a bitrate, but there are a lot of changes between max rate mp3 and really low rate settings.which is expected, to make the files even smaller, they have to get rid of more stuff, and at some point that will include perfectly audible signals.
by default what is assumed to be inaudible to us humans will be removed. you can think of very quiet sounds recorded at the same time as really loud instruments. nobody is going to notice when the crazy quiet stuff are missing. but this system is dynamic, when the only sounds in music are quiet sounds, if I start removing those, you will notice. and the opposite is true, when the music is blasting non stop close to the maximum level, even someone talking at moderate volume might not always be audible to you and we could probably remove that information from the track without you noticing. this is based on auditory masking and it's both simple and not so simple depending how deep down the rabbit hole you're willing to go(and same thing with lossy codecs by the way, some aspects are dead basic, others are advanced math). basically, our hearing is far from perfect, so our music files don't really need to be for us to feel like it is. ^_^
I happened to share this link not long ago. that cool guy explains very intuitively one of the core principles behind lossy encoding


so to answer you question, sort of, stuff may be removed at any frequencies.

 

[Dieucocto]

the low bitrate mp3 ends up getting rid of the upper frequencies above 16khz, but not at 320kbps. it can be confusing because the format keeps the same name and we only move a slider to set a bitrate, but there are a lot of changes between max rate mp3 and really low rate settings.which is expected, to make the files even smaller, they have to get rid of more stuff, and at some point that will include perfectly audible signals.
by default what is assumed to be inaudible to us humans will be removed. you can think of very quiet sounds recorded at the same time as really loud instruments. nobody is going to notice when the crazy quiet stuff are missing. but this system is dynamic, when the only sounds in music are quiet sounds, if I start removing those, you will notice. and the opposite is true, when the music is blasting non stop close to the maximum level, even someone talking at moderate volume might not always be audible to you and we could probably remove that information from the track without you noticing. this is based on auditory masking and it's both simple and not so simple depending how deep down the rabbit hole you're willing to go(and same thing with lossy codecs by the way, some aspects are dead basic, others are advanced math). basically, our hearing is far from perfect, so our music files don't really need to be for us to feel like it is. ^_^
I happened to share this link not long ago. that cool guy explains very intuitively one of the core principles behind lossy encoding


so to answer you question, sort of, stuff may be removed at any frequencies.

Thanks for your answer !

That being said, I didn't said that 320 kbps were getting raid of the upper frequencies above 16 khz. What I meant was that the cut is being made at 20 khz, but we can see, depending on the music, sometimes that between 16 khz and 20 khz there's cleary stuff removed compared to the flac version (not a clear cut, obviously, like on a 128 or 192 kbits mp3). And that's something that we can't see between 0 and 16 khz -- like this area of the 320 kbits was truly lossless.

That's why I'm wondering why can we see the loss between 16/20 khz (again I know that on the 320 kbits the cut is at 20khz, but starting at 16khz we can see a difference in the spectrum comparing to the flac, sometimes it's very thin but it's there), but nothing between 0/16, if that doesn't mean this region of the song is not lossless ? That's the thing I'm wondering about : what's being lost between 0/16 khz on a 320 kbits MP3 comparing to a flac, and why can't we see it on the spectrum ? I know it's not audible, but I'd really like to know how the compression is being made on this area of a song

 

[castleofargh]

Thanks for your answer !

That being said, I didn't said that 320 kbps were getting raid of the upper frequencies above 16 khz. What I meant was that the cut is being made at 20 khz, but we can see, depending on the music, sometimes that between 16 khz and 20 khz there's cleary stuff removed compared to the flac version (not a clear cut, obviously, like on a 128 or 192 kbits mp3). And that's something that we can't see between 0 and 16 khz -- like this area of the 320 kbits was truly lossless.

That's why I'm wondering why can we see the loss between 16/20 khz (again I know that on the 320 kbits the cut is at 20khz, but starting at 16khz we can see a difference in the spectrum comparing to the flac, sometimes it's very thin but it's there), but nothing between 0/16, if that doesn't mean this region of the song is not lossless ? That's the thing I'm wondering about : what's being lost between 0/16 khz on a 320 kbits MP3 comparing to a flac, and why can't we see it on the spectrum ? I know it's not audible, but I'd really like to know how the compression is being made on this area of a song

nothing lossless anywhere. that's just not how mp3/AAC work.
based on usual hearing, it's not a big deal to remove upper frequencies more systematically(as in, without paying as much attention to masking from other sounds in the track as they do at lower freqs). but again, data can be removed/simplified at any frequency based on the principle of auditory masking, and will be! you can think of it as trimming the quieter sounds out of the track, or as increasing the quantization noise because the algorithm expects some loud signals to effectively mask that higher noise when hearing it. it's the same thing from a different point of view. and when the track gets quieter in that frequency range(or fairly close), then the encoding wont let as much noise being added by quantization(and wont save as much space in the file) from fear that it would be noticeable. it's a dynamic process. how much is lost depends on a psychoacoustic model after all.
as to why a spectrum doesn't make it obvious that there is data loss in the music frequency range? well loss of data doesn't mean that you have to lower the amplitude of the music. for example when you turn a 24bit .wav into a 16bit .wav file, would you expect huge changes in the spectrum? yet a lot of data is removed making the file a good deal smaller. bad analogy considering that again MP3 isn't PCM, but maybe that can give you the general idea as to why a spectrum might not be the best tool to see those variations from lossy encoding(even less so if changes by 10dB hardly show on the graph because of the usual scaling settings).

 

[Dieucocto]

nothing lossless anywhere. that's just not how mp3/AAC work.
based on usual hearing, it's not a big deal to remove upper frequencies more systematically(as in, without paying as much attention to masking from other sounds in the track as they do at lower freqs). but again, data can be removed/simplified at any frequency based on the principle of auditory masking, and will be! you can think of it as trimming the quieter sounds out of the track, or as increasing the quantization noise because the algorithm expects some loud signals to effectively mask that higher noise when hearing it. it's the same thing from a different point of view. and when the track gets quieter in that frequency range(or fairly close), then the encoding wont let as much noise being added by quantization(and wont save as much space in the file) from fear that it would be noticeable. it's a dynamic process. how much is lost depends on a psychoacoustic model after all.
as to why a spectrum doesn't make it obvious that there is data loss in the music frequency range? well loss of data doesn't mean that you have to lower the amplitude of the music. for example when you turn a 24bit .wav into a 16bit .wav file, would you expect huge changes in the spectrum? yet a lot of data is removed making the file a good deal smaller. bad analogy considering that again MP3 isn't PCM, but maybe that can give you the general idea as to why a spectrum might not be the best tool to see those variations from lossy encoding(even less so if changes by 10dB hardly show on the graph because of the usual scaling settings).

OK, gotcha ! Thanks !

 

[5twnr]

Something I've noticed - I can hear a difference between 320kbps and V0 VBR pretty consistently. It's like the v0 keeps more echoes and reverbs, while making the bass fuller. There's also better separation of instruments. I noticed this while comparing songs from a Brandy album and songs a Mariah Carey album (to their own albums, not to each other). Compared to FLAC, the v0 sounds more like the FLAC than the 320.

My question is - has anyone else noticed this issue? Are the 320kbps mp3s just encoded poorly? Spectrum analysis shows that they look as expected, but they just sound bad. Now I can't un-notice it. I can provide samples via PM if anyone's interested.

 

[bigshot]

Are you doing level matched, direct A/B switched blind tests?

 

[Brooko]

Something I've noticed - I can hear a difference between 320kbps and V0 VBR pretty consistently. It's like the v0 keeps more echoes and reverbs, while making the bass fuller. There's also better separation of instruments. I noticed this while comparing songs from a Brandy album and songs a Mariah Carey album (to their own albums, not to each other). Compared to FLAC, the v0 sounds more like the FLAC than the 320.

My question is - has anyone else noticed this issue? Are the 320kbps mp3s just encoded poorly? Spectrum analysis shows that they look as expected, but they just sound bad. Now I can't un-notice it. I can provide samples via PM if anyone's interested.

Were both encoded from the same master?

 

[castleofargh]

Something I've noticed - I can hear a difference between 320kbps and V0 VBR pretty consistently. It's like the v0 keeps more echoes and reverbs, while making the bass fuller. There's also better separation of instruments. I noticed this while comparing songs from a Brandy album and songs a Mariah Carey album (to their own albums, not to each other). Compared to FLAC, the v0 sounds more like the FLAC than the 320.

My question is - has anyone else noticed this issue? Are the 320kbps mp3s just encoded poorly? Spectrum analysis shows that they look as expected, but they just sound bad. Now I can't un-notice it. I can provide samples via PM if anyone's interested.

given how people have a really hard time passing a listening test with either of those against lossless, it seems a little strange that you would be able to notice not one but several differences between 2 high bitrate MP3. I'd look up the conversion settings, or question my testing method if I was you.
with that said, on a purely subjective gut feeling level that I was never ever able to confirm under test, I also somehow feel like I prefer VBR. anytime I think I'm hearing a difference, I test for it and fail. but the gut feeling is there anyway ^_^.

 

[5twnr]

I converted the v0 from FLAC source, so on that front I know where the master came from. However with the 320 I'm not sure where it came from, so I only analyzed with Spek to check that the frequencies aren't cut off. To test more fully I should convert some 320s from the flac.

I didn't do a double blind test on v0 vs FLAC, but I can definitely tell the difference between the 320 and FLAC/v0. I also did a blind test for my roommate with the v0 vs 320, and they could tell the difference (they preferred the v0 too) so I guess it's not just me. Just wondering if others have noticed the same.

 

[StandsOnFeet]

Get Foobar 2000 and install the ABX plugin for it. That's the easiest way to see if you can really hear a difference. The test program does everything you want a test program to do, and it gives you a report at the end with the probability that you were guessing.

 

[danadam]

However with the 320 I'm not sure where it came from

I wouldn't look for an answer further than that.

Why isn't it just obvious to everyone that such comparisons, with files from unknown sources, are pointless?

 

[Brooko]

I converted the v0 from FLAC source, so on that front I know where the master came from. However with the 320 I'm not sure where it came from, so I only analyzed with Spek to check that the frequencies aren't cut off. To test more fully I should convert some 320s from the flac.

I didn't do a double blind test on v0 vs FLAC, but I can definitely tell the difference between the 320 and FLAC/v0. I also did a blind test for my roommate with the v0 vs 320, and they could tell the difference (they preferred the v0 too) so I guess it's not just me. Just wondering if others have noticed the same.

Read my sig (the quote ) Expectation bias can be a real issue. Yep - we can all easily tell them apart - when:

1. Its likely the the 2 tracks were completely different masters
2. You didn't blind test - so its also unlikely you volume matched (and "by ear" is not good enough)
3. There was no control in your test, and no repetition to ensure you weren't purely guessing.

Here - tools are free - only costs you time. Results will surprise you

https://www.head-fi.org/threads/set...-guide-to-ripping-tagging-transcoding.655879/

 

[5twnr]

Read my sig (the quote ) Expectation bias can be a real issue. Yep - we can all easily tell them apart - when:

1. Its likely the the 2 tracks were completely different masters
2. You didn't blind test - so its also unlikely you volume matched (and "by ear" is not good enough)
3. There was no control in your test, and no repetition to ensure you weren't purely guessing.

Here - tools are free - only costs you time. Results will surprise you

https://www.head-fi.org/threads/set...-guide-to-ripping-tagging-transcoding.655879/

Are there mac instructions? Can't use most of these tools.

Yeah could be different masters. I'm down to try it again with my own conversions.

 

[bigshot]

The difference isn't likely at all to be from the codec if you are using high data rates. I've done a lot of tests and I've found that MP3 LAME 320 and AAC 256 is perfectly transparent. For most music, lower rates are transparent too.