[music-dsp] Simple Speech Compression

[Joshua Scholar]

----- Original Message ----- 
From: "Tony Robinson" <tony at tonyRobinson.com>
To: <music-dsp at ceait.calarts.edu>
Sent: Monday, August 08, 2005 11:42 PM
Subject: Re: [music-dsp] Simple Speech Compression


> Oh good - I'm glad my post showed up.  Sorry if I'm not following this
> as quickly as others.
>
> > From: "Joshua Scholar" <joshscholar at yahoo.com>
> >
> > Perhaps other compressors have 8 bit modes that do the right thing, but
> > other than Shorten NONE of them do the right thing when handed 16 bit
data
> > that just happens to be mod 256.
> ...
> > I don't see why Tony  assumed that everyone else added the same
> > optimizations that he did.  By the way, shorten beats the more modern
> > losseless compressors in other fringe conditions as well, such as tracks
> > with long silences.
>
> A few reasons - the original quote was "no lossless compressor is going
> to make it smaller than 8 bits per sample" and I was trying to say that
> general lossless compressors (gzip, zip, etc) will pick up on the
> redundancy.  Lossless audio compressors can also make good lossy
> compressors, shorten does this by discarding information such that the
> "noise" added is a fixed level below the signal at each frequency -
> which is a good first order approximation perceptually and can be simply
> implemented by zeroing the bottom bits of the prediction residual.  As I
> wanted good tandem usage this had to be detected.  Also, some hardware
> (I think it was the codec in the first SPARCstations) zeroed some bits
> under some conditions.  Also, I made the source code available from the
> very start, although I haven't looked to see what use other lossless
> codecs have made from these ideas - I hope to get back to it one day.
>

"Lossless audio compressors can also make good lossy compressors,"

No they don't, they make extremely naive lossy audio compressors.  They
manage to keep coding noise localized in time but completely fail to make it
localized in frequency and so color the timbre of the sound with noise
that's very audible to a trained ear.  They also tend to make percussive
sounds noisy in a way that that smears the stereo image.

I've seen better and worse ways to turn lossless compressors into lossy
ones, but none that are good.  In fact the least horrible lossy/lossless
compression I know of is to simple compress to some number of bits other
than 16.  14 bit sound can be quite listenable, infinitely better than sound
which pulses white noise whenever there's the slightest percussive note or
the smallest sibilant.