readme

16kb/s Low Delay CELP 算法

This program implements the 16 kb/s Low-Delay CELP algorithm.
(ITU-T Recommendation G.728)

Original version by Alex Zatsman, Analog Devices, 1993.
Revision by Mike Concannon (Columbia University, NY, NY)  Summer 1994

Disclaimer
----------
This software carries no warranty, expressed or implied.  The user
assumes all risks, known or unknown, direct or indirect, which involve
this software in any way.

-----------------------------------------------------------------------------
About version 1.0 (by Alex Zatsman)

This original version was intended as a demo for EZ-LAB with ADSD21020
(by Analog Devices, Inc.), using the Analog Devides g21k C compiler (port of
gcc to 21k).

It ran on Sparc as encoder (ccelp) and decoder (dcelp). These must be
built by GNU C compiler (gcc) because the source is written in GNU C
extension. In particular, the "{( ... )}" construct is used in several
places (fast.h)

The Sparc version read inputs from Sun audio files (8-bit u-low
encoded). It wrote out the indices into a file using 2 bytes for each
10-bit index. Thus the reduction in file size is 2.5x (instead of 4x
for 8-bit source).

The version 1.0 program has NOT been tested with the official ITU-T test data!

-----------------------------------------------------------------------------
About version 2.0 (based on version 1.0 and modified by Mike Concannon)

All stuff about the sparc audio files have been removed.

The speech files should be 2-byte per sample and contain no header.
The input can use the full 16 bit range.

The bitstreams are written as 2-bytes per 10-bit codeword.

To create, use make all

The command 'dcelp' invokes the decoder + postfilter.  The postfilter may
be turned off by changing the variable 'postfiltering_p' in dmain.c from
1 to 0.
The makefile generates a dcelpnpf version which has the postfilter
turned off.

The postfilter relies on the coefficients (a10 and k10) calculated in the 
durbin recursion (adapters.c) but is otherwise contained wholly in
postfil.c.

To use,
Encoder:
ccelp <input speech file> <output bitstream>

Decoder:
dcelp <input bitstream> <output speech file>

The encoder produced bitstreams match all the test vectors exactly while
the decoder + postfilter output speech files match to between 40 
and 80 dB SNR with the test vector outputs.

See file tv.out
This was generated with the script runtestv, which can only be run
if the testvectors and testprograms are available

VERIFICATION

This code was test on SUN and SGI platforms using the gcc and cc compilers.
In the tdata directory there are a few test files.

f17.in	- inputfile
f17.bit - bitstream file
f17.outnpf - output file (no postfiltering)
f17.outpf - postfiltered outputfile

If you get sampled data files that are not identical, try to run an
snr measure on it, and see if you are close. Any SNR > 40 dB would make
it very likely that what you have is fine.


		List of Sources:

adapters.c:	
	Synthesis Filter Adapter,
	Perceptual Weighting Filter Adapter
	Gain Adapter.

cmain.c:	Main routine for coder.
	
codebook.c:	Function for Codebook Search and Access.

data.c:
		Codebook (shapes and gains)
		Hybrid Window Coefficients for:
			Gain Adapter
			Perceptual Weighting Filter Adapter
			Synthesis Filter Adapter	

dmain.c:	Main routine for decoder.

filters.c:	
		Perceptual Weighting Filter
		Gain Predictor
		Synthesis Filter

global.c:	Variables common for coder and decoder.

io-sparc.c:	Non-21020 (file) I/O routines. Actually NOT specific
		to Sparc machine.

postfil.c:	Postfilter and its Adapter.


common.h:	Common definitions used in most files

data.h:		
		Derivatives from gain codebook.
		Coefficients for low-pass IIR filter (in postfilter)

fast.h:		Some C macros which translate into more efficient code
		for ADSP21000

parm.h:		Various parameters of LD-CELP (vector size, frame
		size, etc)

prototyp.h:	Most function prototypes.

qsize.h:	Size of the I/O buffer (queue). It is separated
		because it is included in some .asm files which don't
		like more sophisticated C macros.