Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!warwick!pipex!uunet!europa.eng.gtefsd.com!darwin.sura.net!newsserver.jvnc.net!newsserver.technet.sg!solomon.technet.sg!alexlee
From: alexlee@solomon.technet.sg (Dr Alex Lee)
Subject: Re: speech compression
Message-ID: <C9vJF1.M87@newsserver.technet.sg>
Sender: news@newsserver.technet.sg
Nntp-Posting-Host: solomon.technet.sg
X-Newsreader: TIN [version 1.2 PL0]
References: <20uorg$6ek@vtserf.cc.vt.edu> <2176jkINNfj6@master.cs.rose-hulman.edu>
Date: Fri, 9 Jul 1993 01:42:36 GMT
Lines: 28


: Has anyone ported the CELP(Joe Campbell) code to any other workstation besides SPARC?  I  
: can't get it to work on a Sun 3/60, NeXT, HP PA-Risc, HP 9000-345, or a linux box.  Any  
: clues?

I have the CELP code compiled and in working condition on a DecStation
running under Ultrix.  I also have a PC version compiled with Borland
C and a faster version compiled with gcc (djgpp).

The problem may be caused by a bug in the code in celp.c:

The array sold[] was NOT initialized to 0.0, and so on systems which
do not automatically initialize all uninitialized variables to zeros,
a lot of probblems, such as overflows, occur.

Although the code I compiled seems to work, the quality of the
transcoded speech is quite poor.  It's intelligible, and I can
almost recognize my voice, but the quality is far poorer than
that acheived by the GSM trancoder.  On my 486/50Mhz PC, the gcc
version runs at 15 times slower than real time in compression, and
about 2 times real time in decompression.

Some one in an earlier post said that the quality he/she achieved
using CELp on a Sun was quite good.  I would appreciate some samples
the writer compressed for testing.  Please send the hexadecimal stream
files.

Alex Lee
