🚀 go-pugleaf

ja51359@uxa.cso.uiuc.edu (axelrod) writes:
>	I'm looking for a audio compression algorithm that will result
>in telephone quality reproduction.  I.E.  4Khz bandwidth, limited dynamic
>range, average S/N ratio.
>	I'm already familiar with using delta-fibonacci, delta-huffman
>techniques, but I'm looking for a more lossy algorithm that will give
>better compression results, more like on the order of 8:1 with 8-bit
>samples.
>	How is the quality of CELP compression?  I heard voices end up
>sounding robotic.  I'd like something that sounds natural.

Our version of CELP, QCELP, sounds quite decent.  If things aren't
tuned just right, voices can get a "sharpness" to them.  To my ear it
sounds superior to standard telephone, and those I've called have been
unable to tell when I'm calling with the desk phone and with the
cellular phone unless we introduced plenty of noise into the system
(at which time the voice starts sounding somewhat "bubbly" as the
noise overwhelms our error correction).

We output 192 bits per 20 millisecond frame which works out to 1200
bytes per second, or 4.3 megabytes per hour of speech.  In addition,
we do voice activity detection and can produce half, quarter, and
eighth rate frames.  The voice activity factor of standard speech
works out to about 0.6 with this method, which means that the
resulting data is only 60% of the size of that where we force it to
stay in full rate mode, which gives about 2.6 megs for an hour of
speech, or 720 bytes per second.

Given your sampling rate of 8000 Hz with 8 bit samples, that would be
28.8 megs for an hour of speech, so we're doing around 7:1 without
even bothering with voice activity, about 11:1 with it, including
error correction.

We're doing all this in an ASIC, but it demonstrates that it's
possible to get what you want with a version of CELP.  At least it
might be worth looking into.
-- 
I never made a mistake in my life.  I thought I did once, but I was wrong.