xspectrum.1

speech signal process tools

.\" Copyright (c) 1987-1990 AT&T, Inc.
.\" Copyright (c) 1986-1990 Entropic Speech, Inc.
.\" Copyright (c) 1990-1997 Entropic Research Laboratory, Inc. All rights reserved.
.\" @(#)xspectrum.1	1.12 10/3/97 ATT/ESI/ERL
.ds ]W (c) 1997 Entropic Research Laboratory, Inc.
.TH XSPECTRUM 1\-ESPS 10/3/97
.if t .ds - \(em\h'-0.5m'\(em
.if n .ds - ---
.if n .ds lq \&"\"
.if t .ds lq \&``
.if n .ds rq \&"\"
.if t .ds rq \&''
.de qI
.ie '\\$2'' \&\\*(lq\fI\\$1\fP\\*(rq
.el \&\\*(lq\fI\\$1\fP \\$2\\*(rq
..
.ds wD \fIwaves+ Manual\fP
.ds wR \fIwaves+ Reference\fP
.ds wT \fIIntroducing waves+\fP
.ad l
.SH "NAME"
xspectrum \- spectrum estimation attachment for xwaves
.SH SYNOPSIS
.B "xspectrum"
[
.BI \-c " host_X_registry"
] [
.BI \-n " host_name"
] [
.BI \-w " wave_pro"
]
.SH DESCRIPTION
.PP
.I xspectrum
is an
.I attachment
for the program
.IR xwaves (1\-ESPS):
a program that runs in close cooperation with
.I xwaves
and provides additional capabilities.
.I xspectrum
facilitates interactive power-spectrum analysis of data displayed by
.I xwaves.
This manual entry briefly describes
.I xspectrum
but is not intended to be a complete
.I xspectrum
manual.
For full details, see the chapter
.qI xspectrum
in \*(wD
and the chapters
.qI xspectrum "reference"
and
.qI xspectrum "quick reference"
in \*(wR.
Also see \*(wD and \*(wR for information about
.I xwaves.
In \*(wT, a hands-on tutorial introduction to
.I xwaves,
there is a demonstration of the use of
.IR xspectrum :
see the section \*(lqSpectrum estimation: \fIxspectrum\fP\*(rq
in the chapter \*(lqSpectrum estimation, adding labels\*(rq.
.SS Introduction
.PP
An attachment is a program that extends the capabilities of
.I xwaves
while running as a separate UNIX process.
Attachments exchange information with
.I xwaves
by communicating through the X server, using a communication protocol
that is compatible with Tcl/Tk.
.PP
.I xspectrum,
an
.I xwaves
attachment, is a general-purpose frequency-spectrum estimation program.
.PP
.I xspectrum
will compute and display power-spectrum estimates from segments of
sampled data in
.I xwaves
waveform display windows.
Individual spectra displayed by
.I xspectrum
can be compared by overlaying them on a common plot.
The spectrum analysis method and parameters can be varied
by entering values in the
.I xspectrum
control panel.
If a linear-prediction (maximum-entropy) spectrum-analysis method
is used,
.I xspectrum
also supports inverse filtering of the selected data
and formant/bandwidth estimation.
.I xspectrum
will also display
\*(lqspectral slices\*(rq\*-single-frame power spectra\*-from data in
.I xwaves
spectrogram windows.
.SS "Starting xspectrum"
.PP
You can start
.I xspectrum
in various ways.
The commonest way is to click the
.B xspectrum
button in the
.B Attach:
item of the main control panel.
Another way is to issue the
.I xwaves
command
.I attach.
The names that appear in the
.B Attach:
item are determined by the global variable
.I attachments,
which can be specified in the
.I .wave_pro
file.
If
.I xspectrum
has several distinct names (e.g. via symbolic links),
each one can be bound to a separate
.B Attach:
button and invoked as a separate attachment.
Thus multiple simultaneous invocations of
.I xspectrum
are possible.
.SS "Creating spectra"
.PP
When
.I xspectrum
is attached to
.I xwaves,
the menu item
.B "xspectrum"
is added to the
.I xwaves
waveform and spectrogram menus.
Spectral analysis is performed by selecting the menu item.
This operation will only succeed if performed
either in a waveform window displaying a one-dimensional signal
or in a spectrogram window.
If the operation is performed in a waveform window,
the spectrum is computed
either on the data in the currently marked segment
or on the data in an analysis window
centered at the current cursor position\*-which of these occurs
depends on settings in the
.I xspectrum
control panel (see below).
After the spectrum is computed and displayed,
the left and right
.I xwaves
markers show the analysis window limits.
The computed values can be written to an ASCII output file.
.PP
When one of the analysis parameters is changed in the
.I xspectrum
control panel, the active spectra are recomputed using the new
parameters.
.PP
When
.B xspectrum
is selected from a spectrogram window menu, no computation is done;
the analysis frame under the cursor is displayed in the
.I xspectrum
window.
Thus the parameters in the control panel have no effect on the image.
.SS Spectrum display window
.PP
Spectral slices and spectra computed by
.I xspectrum
are displayed
in a window popped up by
.I xspectrum.
This display window has a
frequency/amplitude cursor or cross-hair, which can be moved with the mouse,
and an optional reticle.
Moving the frequency cursor causes the frequency cursors in
.I xwaves
spectrogram windows to move in coordination.
Numeric display of frequency and
amplitude are available in the upper-center part of the window.
The time corresponding to the center of the analysis window is printed in
the display's frame.
The cursor can be left at a particular frequency
by removing the mouse pointer from the window with the middle button
depressed.
.PP
The
.I xspectrum
display can be zoomed in frequency and amplitude to a marked region.
Vertical markers are moved using the \*(lqup/down move\*(rq paradigm
(press, drag, release) with the left mouse button.
Horizontal markers are moved in the same way with the shift key depressed.
By default, if a non-null region is delimited with the
horizontal or vertical markers,
zooming takes place immediately to that region.
If the left and right
(top and bottom) markers are coincident, the full frequency
(amplitude) range will be redisplayed.
The right-button menu permits enabling/disabling this feature.
.PP
The right mouse button brings up a menu with options for
saving the current spectrum as a reference,
clearing saved (or all) spectra,
enabling/disabling the zooming feature,
invoking inverse filtering,
and sending
.I xspectrum
plots to a printer or a plot file.
Up to four spectra can be saved as references.
The reference spectra and the current new spectrum
are all displayed in different colors,
with separate cursors and digital readouts.
.SS Control panel
.PP
Some panel items allow you to select among wired-in alternatives.
Other items permit you to enter a numerical value from the keyboard.
Mousing menu selection items with the right mouse
will display the available alternatives.
Menu selection items are indicated with the
standard down-pointer symbol; the current selection is always
displayed.
For numerical inputs, just mouse the item and enter the
value from the keyboard, followed by a
.I RETURN.
.TP
.B "Analysis Type:"
This item selects the spectrum-analysis method.
The available methods are two Fourier-transform methods,
a cepstrally smoothed FFT method, and six maximum-entropy methods.
The two Fourier methods are log magnitude
spectrum using a radix-2 FFT
.RB ( DFT )
and a
.if n O(n^2)
.if t .IR O ( n \v'-0.4'\s-3\&2\s+3\v'+0.4')
discrete Fourier transform
.RB ( DFTR )
where the transform size is exactly the analysis window size.
The cepstrally smoothed method
.RB ( CEPST )
expands on a suggestion originally provided by the MIT-LCS group;
high-pass and low-pass liftering are provided,
with control over the liftering transition region.
(See
.BR "Cep. cut (sec):" ,
.BR "Cep. trans:" ,
and
.B "Liftering:"
below.)
The maximum-entropy methods
are all based on the LPC-style analysis methods supported by the functions
.IR compute_rc (3\-ESPS)
and
.IR refcof (1\-ESPS),
namely autocorrelation method
.RB ( AUTOC ),
covariance method
.RB ( COV ),
Burg method
.RB ( BURG ),
modified Burg method
.RB ( MBURG ),
fast modified Burg method
.RB ( FBURG ),
structured covariance
.RB ( STRCOV
and
.BR STRCOV1 ),
and vector Burg
.RB ( VBURG ),
a fast approximation to structured covariance.
Of the two structured covariance methods, the first
.RB ( STRCOV )
is considerably faster and better behaved.
In each of these LPC cases,
a log-magnitude (maximum-entropy) spectrum is created
via a DFT of the filter coefficients.
For more information on the methods, see
.IR refcof (1\-ESPS),
.IR compute_rc (3\-ESPS),
and the man pages for functions that are mentioned there.
.IP
All computations are performed using floating point arithmetic.
The minimum size FFT used is 512 points, zero-padded as necessary.
For the
.B "DFT"
function, the maximum size FFT is essentially unlimited,
though there is a soft limit of 2097152
which can be changed by resetting the variable
.I xspectrum_max_fft_size
using the
.I xwaves
command
.I send.
The maximum DFT size for the
.B "DFTR"
function is 1025 points.
The latter limit is imposed
to prevent accidental computation of DFTs
that would take impractically long to compute.
.TP
.B "order:"
This sets the order for LPC analysis
underlying the maximum-entropy methods
(if specified by
.BR "Analysis type" ).
The maximum order available is 200
(but may be less depending on the startup profile).
.TP
.B "Window type:"
This selects the time weighting function to be applied before analysis.
The four supported types are
.BR rectangular ,
.BR Hamming ,
.BR Cos^4 ,
and
.BR Hanning .
The default
weighting type is specific to the analysis method.
In particular,
.B "Hanning"
is the default for
.BR DFT ,
.BR DFTR ,
.BR AUTOC ,
and
.BR COV ,
while
.B "rectangular"
is the default for all of the other (maximum-entropy) methods.
Whenever the existing default window method is changed via this item,
the new method is retained by
.I xspectrum
as a new default for the current analysis method
(i.e., for the
.B "Analysis Type:"
in effect when
.B "Window type"
is changed).
.TP
.B "size (sec):"
This is the duration of the analysis window
for input when the
.B "Window limits from:"
item (see below)
is set to
.BR "Cursor +\- size/2" .
.TP
.B "Window limits from:"
This determines how the size and location
of the analysis window are determined.
When
.B "Cursor +\- size/2"
is selected, the window is centered at the cursor,
and its total duration is given by the
.B "size (sec):"
item.
When
.B "Markers"
is selected, the window limits
are determined by the left and right marker positions
in the \fIxwaves\fP windows when the
.B "xspectrum"
item is selected from an
.I xwaves
menu.
The size of the FFT used to perform the computations
is expanded (in powers of 2) as necessary (up to
.IR max_fft_size )
to accommodate the data.
.I max_fft_size
defaults to 2097152, but can be increased arbitrarily in the
.I .wave_pro
file or via the
.I xwaves
command
.I send.
.TP
.B "Preemphasis coeff:"
The coefficient
.I a
of the filter
.if n \{\
.ne 2
H(z) = 1\ -\ az\x'-1'\v'-1'-1\v'+1'.
\}
.if t \{\
.IR H ( z )
=
.RI "1\ \-\ " az \v'-0.4'\s-3\-1\s+3\v'+0.4'.
\}
This 1st-order prefilter is applied to all signals
before spectrum computation.
When this preemphasis is applied, one
extra sample is used from the input sequence to initialize the filter
memory and maintain the requested window size.
.TP
.BR "Inverse Filter Intvl. (sec):" "\ \ \ \-\ \ \ " "Integration Coeff:"
These affect inverse filtering.
When any of the LPC (maximum-entropy) analysis methods have been used,
the linear-prediction coefficients
can be used to inverse-filter the original signal,
yielding a residual signal.
This operation can be initiated by selecting the
.B "inverse filter"
option from the menu brought up by pressing the right mouse button
in the spectrum window.
The amount of the original signal (centered on the analysis window)
to be inverse filtered is determined by the
.B "Inverse Filter Intvl. (sec):"