[Flac-dev] floating point

[Brian Willoughby]

I don't think there is a common practice for what you describe.

32-bit float only has 24 bits of accuracy at any one time.  The  
tricky thing is that these 24 bits shift around in precision, based  
upon the value.  Audio is normalized to +/-1, but samples below  
+/-0.5 would have 25 effective bits, and each halving of the sample  
value adds another bit.

You could convert your 32-bit float audio file to 32-bit integer, but  
you would need to scale such that +/-1.0 takes the full range of the  
32-bit integer (keeping in mind that -1.0 is valid but +1.0 is not,  
thanks to twos-complement coding).  there would still be loss  
compared to the 32-bit float original, but only for samples very  
close to 0.0 in absolute value.

One important question here is now to store a valid 32-bit integer  
sound file, and whether this is supported by the native flac command  
line.  I kinda doubt that there is a standard for this.  So you might  
need to write your own tools based on the FLAC libraries.  If you  
were to read a valid 32-bit float audio file and convert each sample  
to 32-bit integer properly, then you should be able to hand off  
buffer of 32-bit integer samples to the FLAC library and create a  
valid flac file.

Another consideration is that flac compresses quiet audio files more  
than loud audio files.  By expanding the 32-bit float to 32-bit  
integer in the fashion I've described - which is the only way to  
preserve the maximum resolution - you will end up with a flac that is  
much larger than the equivalent 24-bit integer file.  I suppose that  
is obvious, but I wanted to point it out anyway.  It should be much  
better with 32-bit integer audio than it would be with pure noise,  
though.

P.S.  DTS Music Disc format put 5.1 surround into a 16-bit WAV.   
Normally, this would not compress very well with flac.  However, the  
DTS guys realized that some users might play this WAV without a DTS  
decoder, and the white noise could destroy sound systems.  So the DTS  
format drops two bits from every sample, making the white noise 12 dB  
quieter.  An interesting side effect of this is that flac is able to  
compress this much more than regular white noise because of the 12 dB  
reduction.  I only mention this because it is instructive about how  
the amplitude of the audio input affects flac's compression performance.

Brian Willoughby
Sound Consulting


On Aug 7, 2009, at 07:42, Didier Dambrin wrote:
I've tried to find info about unofficial 32bit float support in FLAC,  
and found several conversations.
Most of them were talking about a 24bit limit, but from the manual I  
guess that this limitation is gone, as it supports up to 32bit integer.

So my question is, what would be the best way, or what is a common  
way to FLAC-encode floating point audio?

The first idea is obvious, we have a 32bit float, FLAC supports 32bit  
integer, and it's lossless. So I'd just make it encode 32bit float,  
it's lossless afterall.
2 problems, first I would suspect it wouldn't encode very well, since  
the data wouldn't be "audio" anymore, right? It would be like trying  
to FLAC-encode random data.
So I tried FLACing white noise, and comp ratio was pretty poor,  
confirming this. Interestingly, it still did a better job than ZIP!
The second problem would be that no other tool would read them  
correctly anyway.

The second idea is to truncate & use it as a lossy encoder, which can  
be "audibly lossless" anyway. But what's a common practice here?
I would be tempted to leave 1 bit of headroom above 0dB, or maybe 2  
bits.  Normalizing before encoding could be an option, but the gain  
would then have to be inserted in a tag.
& finally, what bit depth? 24? 25, 26? 32?
I know there's no right answer, just asking if there are common  
practices.