[Flac] Questions: FLAC performance, compression ratio and extra documentation

[Brian Willoughby]

On Nov 12, 2009, at 16:32, Fernando Alberto Marengo Rodriguez wrote:

> I' m studying FLAC performance, and I'd like to know how much  
> compression can be achieved for different audio files.
>
> 1) It seems that for nontonal sound (wideband noise), the  
> compression factor is better than for compound sound (tones +  
> nontonal components), which is typically 2. The reason for this  
> result could be the following: the LPC filter is more suitable for  
> estimating the amplitude spectrum of non tonal information, and do  
> not take into account the phase contribution for each spectral  
> component. Do you agree with me, or do you suggest other reasons  
> for this?

You should also find that lower amplitudes compress better than  
higher amplitudes.  For example, if "live" material with normal  
dynamic range is compressed by FLAC, the ratio will be better than  
the same audio which has been compressed for mastering, which raises  
the levels of all the material in addition to reducing the range  
between quiet and loud.

Your "wideband noise" may be gaussian or uniform distribution, and  
this could greatly affect the compression.  I.e. do not assume that  
all noise is the same.

There are probably other correlations besides just noise vs. tones  
and loud vs. quiet.  It seems like you may want to do additional  
research.


> 2) As far as I know, the compression ratio is defined as
>
> output file size / input file size,
>
> where the output file size includes frame header and other  
> information which do not represent the input samples.

Don't forget that the input file size also includes frame header and  
other information which does not represent the input samples.  This  
input file overhead can be negligible, but is not without some  
consequence.


> My questions are:
>
> a) What percentage of the output sile size is exclusively  
> representing the input data? As this performance may depend on how  
> many tones are present in the input, I'd like to know some average  
> results if possible.


I think you're phrasing the question wrong.  The output size devoted  
to audio data content is variable, so it will not be accurately  
expressed as a percentage.  Instead, there is a fixed amount of  
overhead for a given amount of timeline, which slightly varies  
according to the format chosen (16-bit, 24-bit, etc).  The amount of  
space taken by the audio content is quite variable, and is not  
strictly specified other than algorithmically.  I think you'll have  
to gather statistics rather than look for "specifications," since the  
percentage is a side-effect rather than a precise aspect of the design.


> b) Is redundancy added to the compressed data in order to make FLAC  
> more robust? If this is the case, what's the percentage of these  
> data in the output file?

I believe that you can answer this for yourself by carefully  
examining the details on the FLAC web pages.  The entire format is  
described, down to the bit level.


> 3) Where can I find more documentation about FLAC format and  
> design? Are there any more documents than those in the FLAC web  
> page? It' d be extremely useful not only for me, but also for many  
> people investigating about this interesting open-source format.

I think what you are looking for is really a set of benchmarks that  
report on the effectiveness of the design in reaching its goal.   
There is no more documentation about the FLAC format and design than  
on the web site, because the web site includes everything related to  
format and design.  The numerical factors that you are requesting are  
more about analysis than design, and they are fuzzy values rather  
than fixed aspects of format or design.

Perhaps what you are really looking for is a White Paper which  
describes some of the research which went into FLAC, what sorts of  
sounds were analyzed when developing the algorithm, and what range of  
results were achieved by the resulting software.  I do not recall  
seeing such published information, but perhaps the author, Josh  
Coalson, can elaborate.  I'm sure that he had to test a wide variety  
of sounds, and there are certainly a number of test files in the open  
source suite.  I just don't know how much public documentation there  
is at this high level of analysis.

Brian Willoughby
Sound Consulting