formant.1

speech signal process tools

.\" @(#)formant.1	1.20 20 Sep 1997  ESI
.TH FORMANT 1\-ESPS 20 Sep 1997
.ds ]W "\fI\s+4\ze\h'0.05'e\s-4\v'-0.4m'\fP\(*p\v'0.4m'\ ERL
.if t .ds - \(em\h'-0.5m'\(em
.if n .ds - ---
.SH "NAME"
formant - speech formant and fundamental frequency (pitch) analysis
.SH "SYNOPSIS"
.B formant 
[
.BI \-p " preemphasis"
] [
.BI \-n  " num_formants" 
] [
.BI \-o  " lpc_order"
] [
.BI \-i  " frame_step"
] [
.BI \-w " window_duration"
] [
.BI \-W " window_type"
] [
.BI \-t " lpc_type"
] [
.BI \-F 
] [
.BI \-O " output_path"
] [
.BI \-r " range" 
] [
.BI \-S
] [
.BI \-f " ds_freq"
] [
.BI \-y " f0_max"
] [
.BI \-z " f0_min"
] [
.BI \-N " nom_f0_freq"
] [
.BI \-B " max_buff_bytes"
] [
.BI \-R " maxrms_duration"
] [
.BI \-M " maxrms_value"
] [
.BI \-x " debug_level"
] 
.I " infile"
.SH "DESCRIPTION"
.PP
\fIFormant\fR estimates speech formant trajectories, fundamental
frequency (F0) and related information.  In particular, for each
frame of sampled data, \fIformant\fP estimates formant frequencies,
formant bandwidths, pole frequencies corresponding to linear
predictor coefficients, and voicing information (fundamental
frequency, voiced/unvoiced decision, rms energy, first normalized
autocorrelation, and the first reflection coefficient).
.PP
If only F0 analysis is desired, the new and better F0 estimation
program \fIget_f0\fP(1\-ESPS) should be used, since it is faster, more
accurate and more convenient to use.  \fIGet_f0\fR processes data in
stream mode, and so has no constraints on the length of the input data
sequence (or file).
.PP
Dynamic programming is used to optimize F0 and formant trajectory estimates
by imposing frequency continuity constraints.  The formant frequencies
are selected from candidates proposed by solving for the roots of the
linear predictor polynomial computed periodically from the speech
waveform.  The local costs of all possible mappings of the complex
roots to formant frequencies are computed at each frame based on the
frequencies and bandwidths of the component formants for each mapping.
The cost of connecting each of these mappings with each of the
mappings in the previous frame is then minimized using a modified
Viterbi algorithm.
.PP
The input file
.I infile
is a sampled-data file\*-typically an ESPS FEA_SD file,
though other formats are accpted as well (see
.IR get_feasd_recs (3\-ESPS)).
.I Formant
produces various output files with the same file name body as
\fIinfile\fP (the name body results from removing the last of any
extensions\*-e.g., the name body of "foo.sd" is "foo"), but with
different extensions.  Voicing information is stored in a FEA file
with extension ".f0", formants and bandwidths are stored in a FEA file
with extension ".fb", and pole frequencies are stored in an ASCII file
with extension ".pole".  For details on these and related files that
are relevant for use with \fIxwaves\fP, see "FILE FORMATS", below.
The \fB\-O\fP option permits specification of an alternate output path
in cases where it is undesirable or impossible to write in the input
directory.
.PP
If the sampling frequency of the input speech file is greater than 10
kHz (or the \fB\-f\fP specified value), \fIformant\fP downsamples the
input file to the appropriate sampling rate and saves the results in
an ESPS FEA_SD file with a ".ds" extension.  The input signal, possibly
downsampled, is then high-pass filtered to remove low frequency rumble
(cut-off at approximately 80 Hz), with the result stored in an ESPS FEA_SD
file with a ".hp" extension. The ".hp" file is then used for the
F0 and formant frequency estimates in the manner described above.
.PP
Preemphasis is applied prior to the linear prediction analysis in
order to compensate partially for voice source and radiation
characteristics.  The high-pass filtering is intended to provide a
zero-mean signal for linear prediction analysis when there is a
possibility of residual DC because the real zero of the preemphasis
filter is within the unit circle.  For reliable F0 estimation based on
cross correlations, it is essential that the DC and rumble be removed,
which is another reason for the high pass filtering.
.PP
If ".ds" or ".hp" files already exist in the current directory when 
\fIformant\fP is run, they are used directly and not re-computed.  
This shortcut is intended to save time when analysis conditions 
are varied.  It is, however, somewhat error prone so you should 
be aware of what files are around.   
.SH OPTIONS
.PP
.TP
.BI \-d " debug_level" " \fR[0]\fP"
Values
greater than 0 cause messages to print to stderr.
.TP
.BI \-p " preemphasis" " \fR[.7]\fP"
Specifies the pre-emphasis constant to use before linear predictor
analysis. Possible values range from 0 to 1.
.TP
.BI \-n " num_formants" " \fR[4]\fP"
Specifies the number of formants to attempt to track.  
For adult speakers, 
this
number is normally less than or equal to the sampling frequency in kHz. 
(after down sampling) divided
by 2, and less than or equal to (\fIlpc_order\fP \- 4) / 2.
Currently, a maximum of seven formants are supported.
.TP
.BI \-w " window_duration" " \fR[.049]\fP"
Specifies the length of the data window in seconds over which the
linear predictor analysis takes place.  Note that this default window
size is intended for use with a cos**4 window.  The equivalent length
for a hanning window would be about 25 ms.
.TP
.BI \-W " window_type" " \fR[cos**4]\fP"
Specifies the type of data window to apply to the data prior to
linear predictor analysis. Possible values are 0 (rectangular),
1 (Hamming), 2 (cos**4) and 3 (hanning). 
.TP
.BI \-i " frame_step" " \fR[.01]\fP\"
Specifies the step size in seconds between frames.  This determines
the amount by which the onset of the data window is moved between
analysis frames.
.TP
.BI \-o " lpc_order" " \fR[12]\fP"
Specifies the order of the linear predictor analysis done in each
frame. Maximum order is 30.
.TP
.BI \-f " ds_freq" " \fR[10000]\fP"
Specifies the sampling frequency of the data to be used in the voicing
and formant frequency analysis.  If \fIds_freq\fP is lower than the
input file's sample frequency, the data is down sampled.  Othwewise,
the input sample rate prevails.
.TP
.BI \-N " nom_f1_freq" \fR[500]\fP"
Specifies the nominal value of the first formant frequency.  This
value is used by the program to adjust the nominal values of all other
formants and of the ranges over which the formants are permitted to
exist.  The default value of 500Hz assumes that the vocal tract length
is 17 cm and that the speed of sound is 34000 cm/sec.  Nominal F1
values scale directly with sound velocity and inversely with
vocal-tract length.
.TP
.BI \-t " lpc_type" " \fR[autocorrelation]\fP"
Specifies the linear predictor analysis method.  Possible values
include 0 (autocorrelation) and 1 (stabilized covariance).  If
stabilized covariance method is chosen, however, \fIwindow_duration\fP
(\fB\-w\fP) is set to .025, \fIpreemphasis\fR (\fB\-p\fP) is set to
exp{\-1800*pi/samp_freq), and \fIwindow_type\fR (\fB\-W\fP) is set to
rectangular.
.TP
.BI \-y " f0_max" " \fR[maximum F0 value]\fP"
Specifies the maximum F0 value to search for.  Default is 500 Hz.
.TP
.BI \-y " f0_min" " \fR[minimum F0 value]\fP"
Specifies the minimum F0 value to search for.  Default is 60 Hz.
.TP
.B \-S
Enable the creation of \fISIGnal\fP files for the F0 estimates in
addition to the normal ESPS files.  In previous versions of
\fIformant\fP this was the default, since \fIxwaves\fP required files
with \fISIGnal\fP headers to correctly invoke the special F0 display.
\fIXwaves\fP now develops this display from the ESPS files.  Those who
still desire the old behavior may reenable it with this option.
.TP
.BI \-r " range"
Select a subrange of points to be processed, using the format
.IR start\-end ,
.I start:end 
or
.IR start:+count .
Either the start or the end may be omitted; the beginning or the end
of the file are used if no alternative is specified.  If no range is
specified, the entire input file will be processed.
.IP
If multiple files were specified, the same range from each file is processed.