qlsf.1

speech signal process tools

.\" @(#)qlsf.1	1.9 11/24/87
.TH QLSF 1\-ESPS 11/24/87
.ds ]W "\fI\s+4\ze\h'0.05'e\s-4\v'-0.4m'\fP\(*p\v'0.4m'\ Entropic Speech, Inc.
.ds ]Y "\fBESI INTERNAL\fP
.SH "NAME"
qlsf - quantize LSF-style FEA_ANA files for transmission at 2400 bps.
.SH "SYNOPSIS"
.B qlsf 
[
.BI \-x " debug_level"
] [
.BI \-P " params_file"
] [
.BI \-m " spectral_quant"
] [
.BI \-l " max_steps"
] [
.BI \-p " power_quant"
] [
.BI \-d " duration_quant"
] [
.BI \-v " voiced_steps"
] [
.BI \-u " unvoiced_steps"
]
.I " infile.fea outfile.fea"
.SH "DESCRIPTION"
.PP
.I Qlsf
converts the FEA_ANA file
.I infile.fea 
(with LSF spectral parameters)
into the expanded FEA_ANA file
.I outfile.fea.
In addition to quantizing the pulse powers, pulse durations, and
LSF spectral information, \fIqlsf\fR adds three new integer array fields
\fIraw_power_idx\fR, \fIpulse_len_idx\fR, and \fIspec_param_idx\fR
to the output FEA_ANA file. Descriptions of these auxiliary record
fields are given in \fIaux_ana\fR (5\-ESPS).
.PP
The LSFs are quantized \fIdirectly\fR or as
\fIcenter and offset\fR
frequencies, where the quantization method is specified via the
\fB\-m\fR option. The number of quantization levels per octave 
is also separately specified for voiced and unvoiced speech by the
\fB\-v\fR and the \fB\-u\fR options.
.PP
By default,
pulse power information is represented by the 6-bit power
quantization table \fIr0tab[]\fR  described in [1].
Alternatively, the \fB\-p\fR and \fB\-l\fR options 
can be used to specify a coarser
quantization method. See the descriptions of these
options below for more details.
.PP
By default, \fIqlsf\fR represents the pulse duration information
exactly. A coarser quantization can be specified by use of the \fB\-d\fR
option; see the discussion below for more details.
.PP
The three auxiliary fields contain the table indices that correspond to
the quantized parameter values. The \fIraw_power_idx\fR and
\fIpulse_len_idx\fR fields are the same for both spectral
quantization methods (see\fB \-m\fR below), but the format of the
\fIspec_param_idx\fR field depends on the spectral quantization
method. When the \fIdirect\fR method is used, the \fIspec_param_idx\fR
field contains the table indices corresponding to the quantized LSF
values themselves. When the \fIctr_off\fR method is used, the center
frequency index is coded directly from the table, but the offset
frequency is represented by the difference in table indices between the
center frequency and the higher frequency of the LSF pair.
See [2] for more details on the \fIctr_off\fR quantization scheme.
.PP
If
.I infile.fea
is equal to "\-", 
then standard input is used. If
.I outfile.fea
is equal to "\-", standard output is used.  
.SH OPTIONS
.PP
.TP
.BI \-x " debug_level" " \fR[0]\fP"
.I Debug_levels 
greater than 0 cause messages to print to stderr.
.TP
.BI \-m " spectral_quant"
\fIQlsf\fR currently supports two methods of quantizing
the LSF parameters; this option specifies which
method to use. The two methods are identified by the
strings \fIdirect\fR and \fIctr_off\fR.
The \fIdirect\fR method quantizes each LSF to the closest value
in the table of possible LSFs. (The spacing between LSF table values 
is adjustable by use of the \fB\-v\fR and the \fB\-u\fR options.)
The \fIctr_off\fR method transforms each LSF pair into
a center and offset frequency and then codes each of these in
the table of possible LSF values. (See [2] for more details on the 
center\-offset
computation and quantization.)
.TP
.BI \-v " voiced_steps" " \fR[12]\fP"
An integer value that is
the number of equal logarithmic steps per octave (starting at
400 Hz.) to use for voiced speech. We add two LSF frequencies below
400 Hz. to these: one at 300 Hz and another at 350 Hz.
Reasonable values for \fB\-v\fR range between 6 and 20.
.TP
.BI \-u " unvoiced_steps" " \fR[12]\fP"
An integer value that is
the number of equal logarithmic steps per octave (starting at 400 Hz.)
to use for
unvoiced speech. In addition there is an LSF value at 300 and at
350 Hz.
Reasonable values for \fB\-u\fR range between 6 and 20.
.TP
.BI \-p " power_quant"
As mentioned above, by default
\fIqlsf\fR quantizes the pulse powers with a 6-bit, 1 dB per step
table of pulse power values. One other quantization method is available.
By specifying the sting \fI1.5_dB\fR, a coarser quantization table
is used that has 1.5 dB steps in it. The \fI1.5_dB\fR method also
limits the maximum pulse to pulse variation to 12 dB (8 steps in the table)
by default. The maximum change (in table steps) is selectable via the
\fB\-l\fR option.
.TP
.BI \-l " max_steps"
If the \fB\-p\fR option is used to select the \fI1.5_dB\fR
method of power quantization, the maximum chnage in power from one
pulse to the next may be set with the \fB\-l\fR option. Each step
in the table is 1.5 dB, so setting \fImax_steps\fR to 10 is equivalent
to limiting the maximum pulse to pulse change in power to 15 dB.
If \fB\-l\fR is not specified, the maximum change allowed is 8 steps or 12 dB
for the \fI1.5_dB\fR pulse power quantization method. There is no limit
for the change when the \fI6-bit\fR pulse power quantization method is used.
.TP
.BI \-d " duration_quant"
By default the pulse durations are represented exactly.
A coarser quantization is possible by specifying the string
\fI2_sample\fR. This method uses a 2 sample resolution for pulse durations
less than or equal to 80 samples and a 
4 sample resolution for pulse durations greater than
80 samples. (Note this quantization applies only to voiced speech;
unvoiced speech frame lengths are represented exactly.)
The quantization error accumulated during each segment of voiced speech
is added to the first unvoiced frame that is encountered. 
.SH "ESPS PARAMETERS"
.PP
The parameters \fIspectral_quant\fR, \fIpower_quant\fR, \fIduration_quant\fR,
\fImax_steps\fR,
\fIvoiced_steps\fR, and 
\fIunvoiced_steps\fR are read from the parameter file. Their
possible values are described under the \fB\-m\fR, \fB\-p\fR, \fB\-d\fR,
\fB\-l\fR, \fB\-v\fR, and
the \fB\-u\fR options respectively. Note that the command
line values override any parameter file values. The value of
\fIuvced_frmlen\fR is also read from the parameter file.
.SH ESPS COMMON
.PP
ESPS Common is not used.
.SH ESPS HEADERS
.PP
Values in the header of 
.I outfile.fea
are copied from  the values in the header of 
.I infile.fea.
Three option-related generic header items are added to the output file:
\fIpitch_quant\fR, \fIpower_quant\fR, and \fIlsf_quant\fR.
\fIPitch_quant\fR has the string value \fIexact\fR or \fI2_sample\fR;
\fIpower_quant\fR has the string value \fI6_bits\fR or \fI1.5_dB\fR;
\fIlsf_quant\fR takes on the 
string value specified by the \fB\-m\fR option (either \fIdirect\fR or
\fIctr_off\fR). If \fIpower_quant\fR has the string value \fI1.5_dB\fR,
the generic header item
\fImax_steps\fR is also added. Finally, the parameters \fIunvoiced_steps\fR,
\fIuvced_frmlen\fR, and \fIvoiced_steps\fR are added to the output generic
header.
.SH "FUTURE CHANGES"
.PP
Add the ability to specify the range of possible offset values
for \fIctr_off\fR quantization.
.SH WARNINGS
.PP
\fIQlsf\fR warns and exits if the input file is not a FEA_ANA file, or
if it does not have
\fIspec_rep\fR == LSF, or if the \fIidx\fR fields already exist.
Also, there is no default value for \fB\-m\fR; if no value is specified,
\fIqlsf\fR warns and exits.
LSF filter orders greater than 20 are not supported; \fIqlsf\fR
warns and exits.
.SH "SEE ALSO"
.PP
iqlsf(1\-ESPS), q2400 (1\-ESPS), quant (1\-ESPS), aux_ana(5\-ESPS).
.SH "BUGS"
.PP
None known.
.SH REFERENCES
.PP
[1] Shore, J. E., \fIEntropic 8000 Quantization and Coding Method 
!ros1c!\fR, ETM-S-87-23:js.
.PP
[2] Kang, G. S. and Fransen, L. J., \fILow-Bit Rate Speech Encoders
Based on Line-Spectrum Frequencies (LSFs)\fR, NRL Report 8857.
.SH "AUTHOR"
.PP
Program and manual page by David Burton and Jim Elliott.