Head-Fi.org › Forums › Equipment Forums › Sound Science › lossy audio codec comparison photos and sound clips
New Posts  All Forums:Forum Nav:

lossy audio codec comparison photos and sound clips

post #1 of 9
1/25/12 at 8:44pm
Thread Starter 

This was found on another forum that I can't link to, but I think it deserves to be posted. It gives a visual representation of the information that gets removed from a song when being converted to different formats, and even includes sound files of the removed information. I'm going to go ahead and state that I don't know the specifics of how these photos or files were made, so these are meant to be taken as-is. I've revised the posts a bit to fit into this post. All edits will be bolded by me.

 

This post details the audio inversion methods for the tests.

 

 

 

 

Quote:

- How did you do the spectral subtraction, graphically or by subtracting one wave from the reference, so only the difference remains, then the espectral is calculated.



It isnt that difficult, you just need some physics(interference)/maths

http://upload.wikimedia.org/wikipedia/commons/thumb/0/0f/Interference_of_two_waves.svg/500px-Interference_of_two_waves.svg.png

an audiowave looks like on the left. now, i invert the same wave (in audio its called phase inverting) and i get the left picture when i add those two waves (+1 + -1 = 0).
 
so what ive done is pretty simple. converted a WAV into MP3, putted them together in auditions multitrack view (track#1 WAV, track#2 MP3), adjusted the position, phase inverted the MP3, and so you get the difference between those two tracks. nothing with photoshop.
 
seeya

 

 

Quote:
Here is the source file for these tests. It contains a 4 second clip of a song called "Cool Walk : http://www.mediafire.com/?by0wn328mz7jrgz
 
The audio inversion method definitely yields more tangible results.

Heres the screen shots with the revised inversion method from the samples provided in the previous post.

Only audio differing from original flac shown
AAC 320
http://whatimg.com/images/15685445630934363695_thumb.jpg

MP3 320
http://whatimg.com/images/35971244557369998702_thumb.jpg

V0
http://whatimg.com/images/62742124109024676439_thumb.jpg

V2
http://whatimg.com/images/41551863292994161854_thumb.jpg

OGG 8
http://whatimg.com/images/60508062363690943552_thumb.jpg

Thanks for showing me more accurate ways to conduct these tests folks, appreciated. 

 

And another reproduction by another user

 

 

 

Quote:
Here's some more detailed ones. The song is Awake by Mutiny Within, loud metal with lots of high frequency action.

You have to keep in mind that one of the main principles of lossy encoding is the masking effect, where things that are removed were inaudible when simultaneously sounded against other, louder sounds that were close by in the frequency domain. They become perfectly audible when subtracted from the original soundwave in the invert + mix process. 

I will include the subtracted audio files too once they finish uploading Done.

What was removed from Vorbis q8
http://whatimg.com/images/13889604311667929766.png
http://whatimg.com/images/54105722696290336001.png
What you can hear in this file is mainly remnants from drums, basically sounds like static though. You can hear a little bit of the vocals, mainly the airy part of the voice.
 
http://www.mediafire.com/?7v0im2rt1fwa492

What was removed from MP3 V0
http://whatimg.com/images/53722625175486377326.png
http://whatimg.com/images/40677477407733563510.png
More vocals are audible in the remnants, still behind lots of drum "static." The static is definitely more focused in the higher frequencies compared to Vorbis. There seems to be random noises every once and a while.
 
http://www.mediafire.com/?zz9c8v2zsejmllf

What was removed from AAC q 0.66
http://whatimg.com/images/11889359002591127713.png
http://whatimg.com/images/93073505582632191530.png
This version has the original music most audible. Whether or not that means something is debatable. I can even make out a few snare hits, as well as the "click" of the bass drum, instead of them being masked in the drum static.

 

edit: partially fixed for formatting.


Edited by BobSaysHi - 1/25/12 at 8:47pm
Reply
post #2 of 9
1/25/12 at 9:09pm

None of the images or downloads are working for me.

 

Keep in mind that even if one codec removes more information, it won't necessarily sound worse (or different). Lossy codecs remove what they think we won't hear. If one removes more information, but more of that information is stuff we don't hear or don't notice missing, it's technically doing its job better.

Reply
post #3 of 9
1/25/12 at 9:40pm
Thread Starter 
Quote:
Originally Posted by Head Injury View Post

None of the images or downloads are working for me.

 

Keep in mind that even if one codec removes more information, it won't necessarily sound worse (or different). Lossy codecs remove what they think we won't hear. If one removes more information, but more of that information is stuff we don't hear or don't notice missing, it's technically doing its job better.


dang. I don't really want to upload all those images myself...

 

I'll work on it tomorrow. I was thinking they wouldn't go through, but they looked fine at first.

 

And on the last bit, I'm fully aware. I still think the information has some merit. 

 

Reply
post #4 of 9
1/26/12 at 5:04am

The images don't work because the image host site you are using forbids hot linking, copy the links in your address bar to see the images, if you get a 403 forbidden, don;t use "refresh", use "go". Also I don't know how this is supposed to be relevant, listening to the dicarded parts alone won't give an indication on how efficient the psycho acoustic model is considering it is partly based on masking.

Reply
post #5 of 9
1/26/12 at 6:38am

Those picture posted above basically tell nothing about the quality of a lossy encoding. That's what listening tests are for. For example see listening tests @ HA wiki.


Edited by xnor - 1/26/12 at 6:39am
Reply
post #6 of 9
1/26/12 at 8:15am
There's a very old thread like this one. JPGs were used but it doesn't really matter except on the highest bitrate lossy encodes.
Reply
post #7 of 9
1/26/12 at 11:14am
Quote:
Originally Posted by xnor View Post

Those picture posted above basically tell nothing about the quality of a lossy encoding. That's what listening tests are for.

^ this ^

disclaimer: I cannot see or access any of the pictures or audio files to which the OP linked. My post is based primarily on the older thread to which Lord Crow linked.

A good lossy codec discards information that its model predicts cannot be heard (or, if the bit rate is too low, is least hearable) within the original file. Comparing visual representations of the lossy files and comparing them to the visual representation of the original file tells us nothing about how codecs audibly compare to one another. Likewise, hearing information that was removed from the original file by a lossy codec does not provide insight into whether or not that information was audible when played along with the retained information.
Edited by Jaywalk3r - 1/26/12 at 12:51pm
Reply
post #8 of 9
1/26/12 at 11:33am

Right.  What the residual sounds like and the magnitude of it, may not really represent the quality of the encoding.  You can't really take it out of context for anything more than a comparison out of curiosity.

Reply
post #9 of 9
1/26/12 at 12:49pm
Thread Starter 
Quote:
Originally Posted by mikeaj View Post

Right.  What the residual sounds like and the magnitude of it, may not really represent the quality of the encoding.  You can't really take it out of context for anything more than a comparison out of curiosity.

 

Exactly. These photos or clips aren't meant to be taken at face value. This isn't meant to be super scientific guys. I just saw this a while ago and thought some of you might like to see them too. 

 

Quote:
Originally Posted by Lord Crow View Post

There's a very old thread like this one. JPGs were used but it doesn't really matter except on the highest bitrate lossy encodes.


Another excellent post. The only thing I would say is that those pure tone sweeps aren't necessarily the same as a normal music file. It's not a huge deal though. It's really interesting to see the cut-off points for the different formats.

 

I'll transfer the photos to imgur later today.

Reply
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Sound Science
Head-Fi.org › Forums › Equipment Forums › Sound Science › lossy audio codec comparison photos and sound clips