[FFmpeg-user] Compare quality of 2 audio files

Rodolfo Medina rodolfo.medina at gmail.com
Sat Jul 1 17:13:13 EEST 2017


Harold Tessmann <htessmann at control-tec.com> writes:

> On Thu, Jun 29, 2017 at 12:37 PM, Moritz Barsnick <barsnick at gmx.net> wrote:
>
>> And unlike video, it seems there are few, if any, methods to measure
>> the *perceived* quality of audio. Check here for some ramblings:
>> https://stackoverflow.com/questions/2945531/determining-
>> the-best-audio-quality
>>
>> But I challange you to find an algorithm which can compare two audio
>> tracks. (And you need someone with a good ear to confirm its findings.
>> Or reference material and encodings such as SQAM.) But if you do find
>> it, and it's "free" to use, please implement an ffmpeg filter with it.
>> :-)
>>
>
> To continue along those lines, I don’t think this is even a problem that
> one can precisely define. I will start by acknowledging that there exists a
> boundary where you can define audio quality objectively. Clipping, for
> instance, destroys sound data, and is objectively bad <
> https://en.wikipedia.org/wiki/Loudness_war> (but even then, some musicians
> or sound designers may want that, for industrial or other effects). But
> there is a difference between objective quality and what people want.
> Consider the ability of a display to reproduce an image accurately. You can
> make an objective comparison, but if you go by the display section of an
> electronics store, you’d find out that people tend to like oversaturated,
> objectively worse, color <
> http://www.flatpanelshd.com/focus.php?subaction=showfull&id=1328263571>.
>
> And then you have to consider the environment of your audience, which you
> can’t necessarily control. Audio played in a car has to contend with road
> noise, while pictures on a TV will look significantly different in a bright
> store vs. a customer’s home <
> https://www.cnet.com/news/why-do-plasma-tvs-look-washed-out-in-the-store/>.
>
> And then there’s the question of how much quality your audience can detect.
> The MythBusters did a test with different grades of vodka and determined
> that yes, an expert can taste the difference between high end and cheap
> liquor, but your average person doesn’t have so discriminating a palate <
> http://kwc.org/mythbusters/2006/04/episode_50_bullets_fired_up_vo.html>. So
> even if you did come up with a good metric, perhaps you don’t _want_ to
> check it against an audiophile, or a recording engineer, or somebody who
> has particularly good hearing.


But suppose you have a song, or any piece of music...  a certain particular
precise recording, just that one...  and that you find in internet two mp3
different files of that song, from two different web sites, both same size in
megabytes.  Suppose that the test with Audacity, which we have earlier much
spoken of, reveals a difference between them: this means that a portion of the
original WAV file during the MP3 conversion has been discarded in more quantity
in, say, file1.mp3 than in file2.mp3 (in fact it can be heared).  Or also
suppose that both files are in wav format but still Audacity reveals a
difference between them (because one of them could be - as far as we know - the
result of a previous unknown conversion).  Then I expect if would be
technically possible that we could give an algebraic plus or minus sign to that
discarded portion and tell - objectively and not by a simple personal listening
feeling - which of the two is the one that contains more information and which
less - that's what I would reasonably expect.

Rodolfo



More information about the ffmpeg-user mailing list