Re: [Audacity-devel] Where does LPCM down-mixing code live?

[Martyn S. <mar...@gm...>]

On 16/08/2010 02:44, Bob Ingraham wrote:
>
> On Sun, Aug 15, 2010 at 5:54 PM, Martyn Shaw <mar...@gm...
> <mailto:mar...@gm...>> wrote:
>
>     Convert to float/double before all other operations?  Then you won't
>     get the clipping of adding 16-bit floats together.  Just a thought.
>
> That makes sense, and that is what I am doing.

Good.

> However, after all of the sample addition has been completed, I still
> have to export the resultant downmixed stream back to a 16-but signed
> format, at which point, it will

'may'

clip unless I apply some sort of
> constant scaling factor to scale the waveform back into the 16-bit domain.

True.  We allow that to happen and don't do anything automatically. 
We keep the 'real' values as floats internally but limit to +-1 for 
export / 'View->Show Clipping'.

What should that 'constant scaling factor' be?  You could search for 
the peak and then divide out, taking away user control.  Or use '/2' 
or some other magic number.  We have Effect->Amplify... and are happy 
with that.  You may have a different situation where you want to do 
that automatically.

Dividing by the number of streams will typically give you an 
over-quiet result if the sources are uncorrelated, which most 
real-world sources are.

Apologies if I misled you before.

HTH
Martyn

PS Never multiply the samples together.

>     We do it (eventually) in Mix.cpp MixBuffers li317.
>
> I read through Mix.cpp and saw that sample were simply added as floats
> (with what looked like an optional per-channel gain multiplier.)
> But I assume that the export function must take care of the scaling
> because that is where it is going to have to write-out the resultant
> file as 16-bit signed, 44.1KHz, etc (assuming that is what I tell it to
> export to.)
>
>     HTH
>     Martyn
>
> Thanks!
>
>     On 12/08/2010 03:58, Bob Ingraham wrote:
>
>         Hello,
>         I have been attempting to implement a software-only Linear PCM
>         down-mixing algorithm (i.e., combining multiple LPCM streams
>         (16-bit-signed, 44.1KHz, 2-channel) into a LPCM single stream,)
>         in Java
>         but am having limited success with finding a good algorithm.
>         Lots of Googling and experimentation but mediocre down-mixed
>         results.
>         Apparently, adding LPCM samples together and dividing by the total
>         number of samples added, doesn't cut it...  That's where I
>         started and
>         then went on tweaking from there, but the best result I get is
>         that I
>         can distinguish the individual source streams, but there's so much
>         "static" noise introduced by my algorithm that it renders the result
>         unuseable.
>         So, I've downloaded the Beta source from SVN and am starting to go
>         through it to see if I can find out how the experts do it.
>         Which brings me to my question: Would someone be kind enough to
>         point me
>         to the part of the source tree that deals with down-mixing LPCM
>         streams?
>         Thanks,
>         Bob
>
>
>
>         ------------------------------------------------------------------------------
>         This SF.net email is sponsored by
>
>         Make an app they can't live without
>         Enter the BlackBerry Developer Challenge
>         http://p.sf.net/sfu/RIM-dev2dev
>
>
>
>         _______________________________________________
>         audacity-devel mailing list
>         aud...@li...
>         <mailto:aud...@li...>
>         https://lists.sourceforge.net/lists/listinfo/audacity-devel
>
>