rasta.1

ears-0.32, linux下有用的语音信号处理工具包

.\" A few notes:
.\" i) Don't add extra blank lines - these appear in the output and
.\"    leaves too much space when display on the screen
.\"
.\" This title line needs to be changed to contain your program name,
.\" but the date will be changed automatically by RCS.
.TH rasta 1 "$Date: 1995/07/03 21:16:59 $" ICSI
.\" Don't do anything clever in the section below - it messes up the
.\" whatis database that's accessed using "man -k"
.SH NAME
rasta \- computes plp or rasta-plp features.
.SH SYNOPSIS
.B rasta
.RB "[" options "]"
.SH DESCRIPTION
.I rasta
computes variants of the Perceptual Linear Prediction (PLP)
analysis, including several forms of RASTA processing. It reads
one of several possible input files containing audio waveforms,
and outputs one of several formats of feature vectors, one vector per
frame.  At least one option (described below) is required.
.SH OPTIONS
.TP 1i			\" Indent the paragraph
.B \-O " [FALSE] "
for online processing 
.TP
.B \-d " [FALSE]"
for debug output 
.TP
.B \-M " [FALSE]"
for small constant added to pspectrum 
which masks small problems in recording
as can be seen in the Switchboard corpus
.TP
.B \-F " [FALSE]"
for highpass filter on input (for dc offset) 
.TP
.B \-a " [FALSE]"
for input ascii file.
Default input is binary shorts
.TP
.B \-A " [FALSE]"
for ascii output file.. 
Default output is binary floats.
.TP
.B \-e " [FALSE]"
for ESPS format input files;
overrules other input options.
.TP
.B \-E " [FALSE]"
for ESPS format output files;
overrules other output options.
.TP
.B \-b " [FALSE]"
for MAT format input files;
overrules other input options.
.TP
.B \-B " [FALSE]"
for MAT format output files;
overrules other output options.
.TP
.B \-T " [FALSE]"
swap bytes on input.
This option makes sense only with binary shorts as input.
.TP
.B \-L " [FALSE]"
for log rasta 
.TP
.B \-J " [FALSE]"
for JAH rasta 
.TP
.B \-C " [FALSE]"
for constant JAH.
Default is adapting JAH according to noisepower.
Only use option -C when -J is used
.TP
.B \-R " [FALSE]"
for getting critical band values as output (ASCII)
instead of cepstral coefficients.
This is useful for users who would like to find
their own spectral mapping coefficients.
This overrules the -P option.
.TP
.B \-P " [FALSE]"
for getting cube root compressed and equalized
critical band values as output instead of cepstral coefficients.
.TP
.B \-g " [compute gain]"
don't compute gain.
.TP
.B \-h " [FALSE]"
use stored noise level estimation and RASTA filter
history for initialization (if history file is
available, otherwise use normal initialization).
.TP
.BI \-i " input file" " [stdin]"
.TP
.BI \-o " output file" " [stdout]"
.TP
.BI \-f " mapping file" " [map_weights.dat]"
JAH Rasta mapping coefficients input text file; specifies
a multi-linear regression to reduce the variance introduced by
using different J values.
.TP
.BI \-r " rasta mixing coefficient" " [1.0]"
for partially rasta, partially plp
e.g. -r 1.0 => no mixing, -r 0.8 => 80% rasta 20% PLP
.TP
.BI \-w " analysis window size" " [20]"
(in milliseconds)
.TP
.BI \-W " windowing constant" " [0.540000]"
.TP
.BI \-s " window step size" " [10]"
(in milliseconds)
.TP
.BI \-S " Sampling frequency" " [8000]"
(in Hertz)
.TP
.BI \-l " liftering exponent" " [0.600000]"
.TP
.BI \-p " pole value" " [0.940000]"
.TP
.BI \-m " model order" " [8]"
.TP
.BI \-n " number of output parameters" 
where default is model order plus 1 (log gain) 
and log gain is given first
.TP
.BI \-c " number of crit band filters"
where default depends on sampling freq, but is 17 for 8000 Hz
.TP
.BI \-N " numerator rastafilt file"
unimplemented in release version 2.2
.TP
.BI \-D " denominator rastafilt file"
unimplemented in release version 2.2
.TP
.BI \-j " Constant J" " [1.000000e-06] "
depends on the noise level, ( smaller for more noisy speech).
Only use option -j when -C is used.
.TP
.BI \-H " History filename" " [history.out] "
File containing noise level and filter history for initialization.
.SH EXAMPLES
To compute log RASTA-PLP (which is optimal for the case of convolutional
error, as when the speech input has been run through a linear
time-invariant filter), with ascii input and ESPS-compatible output:

	rasta -a -E -L < infile.asc > esps_outfile

This will use all the defaults, including an assumption of 8 kHz
sampling rate for the input data. Note that the use of ESPS requires
making rasta using the ESPS-based Makefile, and also requires an
ESPS license. It will compute 9 log RASTA cepstral coefficients (8th
order, including a gain term).


To compute 12th order J-RASTA-PLP (which handles additive noise in addition
to the convolutional), using automatic noise estimation, and with
ascii input and output and a 16 kHz sampling rate:

	rasta -a -A -i speechin.asc -o speechout.asc -J -S 16000 -n 12

This will assume the existence of a mapping file
for different J-values in map_weights.dat .

To compute PLP without any RASTA filtering, using binary shorts
at the input and binary floats at the output (and all the other
defaults):

	rasta < speechin > speechout
.SH FILES
The default mapping coefficients input text file is 
.B map_weights.dat .
You can use the -f option to specify any other name.
.SH AUTHOR
Nelson Morgan (original version, after a Fortran PLP by Hynek Hermansky)
\<[e-mail]\>
.br \" How to break lines between authors
Later additions - Grace Tong, Chris Ehrlicher, Brian Kingsbury
\<[e-mail]\>
.SH SEE ALSO
.BR "Hermansky, H., and Morgan, N., ``Rasta Processing of Speech,''"
.ul
IEEE Transactions on Speech and Audio Processing,
special issue on Robust Speech Recognition, vol.2 no. 4, pp. 578-589,
Oct., 1994
.SH BUGS
Version 2.2 doesn't yet implement the -N or -D switches
to specify the rasta filter.  It also doesn't eliminate all problems
in speech recognition yet.