man.diff

来自「speech signal process tools」· DIFF 代码 · 共 247 行

.\" Copyright (c) 1986-1990 Entropic Speech, Inc.
.\" Copyright (c) 1996 Entropic Research Laboratory, Inc. All rights reserved.
.\" @(#)find_ep.1	1.12 4/1/97 ERL
.ds ]W (c) 1996 Entropic Research Laboratory, Inc.
.TH FIND_EP 1\-ESPS 4/1/97
.SH "NAME"
find_ep \- analysis to locate endpoints of a speech segment in a waveform file
.SH "SYNOPSIS"
.B find_ep 
[
.BI \-b " adbits"
] [
.BI \-c " context"
] [
.B \-e
] [
.BI \-f " final_thresh"
] [
.BI \-h " high_thresh"
] [
.BI \-l " low_thresh"
] [
.B \-n 
] [
.BI \-p " start_point"
] [
.BI \-r " start_point"
] [
.BI \-s " silence_req"
] [
.BI \-t " time"
] [
.B \-w 
] [
.BI \-x " debug_level"
] [
.I " infile.sd"
] [
.I " outfile.sd"
]
.SH "DESCRIPTION"
.PP
\fIFind_ep\fR uses thresholds to find isolated words
or connected segments of speech
in a sampled-data file (FEA_SD(5\-\s-1ESPS\s+1))
and sets \fIstart\fR and \fInan\fR in ESPS Common to
point to the beginning and end of the selected utterance. 
These thresholds may be set by the user
(via the \fB\-l\fR, \fB\-h\fR, and \fB\-f\fR options), or
default values may be computed by the program. If
neither \fB\-p\fR (or \fB\-r\fP) nor \fB\-n\fR is given,
\fIfind_ep\fR starts computing statistics 
at the beginning of the sampled-data file.
If a \fIstart_point\fR is given, \fIfind_ep\fR starts there
for determining thresholds, but starts looking for words \fIsilence_req\fR
milliseconds later in the file.
If the \fB\-n\fR option is set,
\fIfind_ep\fR looks in ESPS Common for \fIstart\fR and \fInan\fR
and starts processing the input file (computing thresholds)
50 milliseconds
beyond the point \fIstart + nan\fR.
.PP
In brief, \fIfind_ep\fR determines the endpoints of an utterance
by moving through the file in \fIcontext\fR millisecond chunks
of speech. In each \fIcontext\fR chunk, \fIfind_ep\fR compares
the average adjusted magnitude (AAM) with one of the three thresholds.
(See [1] for a definition of AAM.) Once the AAM has exceeded
the \fIlow_thresh\fR and then the \fIhigh_thresh\fR, and it has
subsequently fallen below the \fIfinal_thresh\fR and remains below
the \fIlow_thresh\fR for \fItime\fR milliseconds,
\fIfind_ep\fR declares that
a utterance has been detected.
.PP
\fIFind_ep\fR implements a heuristic 
method for detecting the beginning and ending points
of speech in low to moderate background noise environments.
It is not perfect, but it is often useful for finding
isolated words or segments of continuous speech.
.PP
By default, \fIfind_ep\fR writes no output file; it simply
updates \fIstart\fR and \fInan\fR in common. If a sampled-data file
containing only the selected points is desired,
the \fB\-w\fR should be used and an output sampled-data file name 
(\fIoutfile.sd\fR) should
be specified on the command line. This will result in a new file that
contains only the points between \fIstart\fR and 
\fIstart\fR + \fInan\fR \-1 
from the input sampled-data file (\fIi.e.\fR 
the points selected by \fIfind_ep\fR).
.PP
If
.I infile.fea
is equal to "\-" and the \fB\-n\fR option is not used, 
then standard input is used. 
.SH OPTIONS
.PP
.TP
.BI \-b " adbits" " \fR[12]\fP"
Set the number of bits that were used by the A/D converter to
record the data. It is important to do this correctly. This number
is used to scale all the thresholds.
If the data was not taken directly from an A/D converter,
try setting this value to the number of bits used in the original
A/D conversion. You may have to adjust this value in order to
get \fIfind_ep\fR to work reliably, however.
.TP
.BI \-c " context" " \fR[10]\fP"
The time interval in milliseconds that is used for computing AAM.
Note: \fIsilence_req\fR and \fItime\fR will be rounded to an even
multiple of \fIcontext\fR.
.TP
.B \-e 
If this option is set,
\fIPossible word at end of file\fR is not considered an error.
A warning message is written to \fIstderr\fR, but \fIstart\fR and \fInan\fR
are set to the start of the word and the end of file respectively.
.TP
.BI \-f " final_thresh"
Average magnitude threshold for the end of a word. The default value
is .8*\fIlow_thresh\fR. Note \fIfinal_thresh\fR must be less than
\fIlow_thresh\fR.
.TP
.BI \-h " high_thresh"
The second threshold that the potential speech utterance must exceed
before a speech utterance is detected. The default value is 
4*\fIlow_thresh\fR.
.TP
.BI \-l " low_thresh"
Initial average magnitude threshold for detecting the start of an
utterance. A default value may be computed by the program from the
statistics of the first \fItime\fR milliseconds of data.
.TP
.B \-n 
Get the next speech utterance after the current range as specified
in ESPS Common.
If \fB\-n\fR is given, \fB\-p\fR and \fB\-r\fR are ignored.
.TP
.BI \-p " start_point" " \fR[1]\fP"
\fB\-p\fP is a synonym for \fB\-r\fP.
.TP
.BI \-r " start_point" " \fR[1]\fP"
The starting point in the sampled-data file may be specified.
Note that this starting point is for computing statistics for threshold 
computation. If \fIlow_thresh\fR is manually set (\fB\-l\fR),
then the starting point is for finding speech \- that is, no data is
used for computing thresholds.
.TP
.BI \-s " silence_req" " \fR[200]\fP\"
This is the time in milliseconds required to mark the end of an utterance,
and it is the time period required between adjacent utterances in a single
sampled-data file.
Values below 150 milliseconds are not recommended,
but the only enforced lower bound is
.I context.
.TP
.BI \-t " time" " \fR[\fIsilence_req\fR \- 50]"
The number of milliseconds of data used to compute thresholds.
This is ignored if \fIlow_thresh\fR is supplied.
.TP
.B \-w 
Write the selected segment of speech to the output sampled-data file
(\fIoutfile.sd\fR).
An output file name needs to be specified only if this option
is set.
.TP
.BI \-x " debug_level" " \fR[0]\fP"
Values greater than 0 cause messages to print to stderr.
.SH "ESPS PARAMETERS"
.PP
The PARAMS file is not processed.
.SH ESPS COMMON
.PP
If no input file name is supplied,
\fIfind_ep\fR looks in ESPS Common for the input file name.
If \fB\-n\fR is specified,
\fIfind_ep\fR checks the input file name with that in Common,
and if they are consistent, \fIfind_ep\fR 
determines the starting point by using the Common values
for \fIstart\fR and \fInan\fR. 
Finally,
the Common values for \fIstart\fR and \fInan\fR are always updated
by \fIfind_ep\fR.
.PP
ESPS Common processing may be disabled by setting the environment variable
USE_ESPS_COMMON to "off".  The default ESPS Common file is .espscom 
in the user's home directory.  This may be overidden by setting
the environment variable ESPSCOM to the desired path.  User feedback of
Common processing is determined by the environment variable ESPS_VERBOSE,
with 0 causing no feedback and increasing levels causing increasingly
detailed feedback.  If ESPS_VERBOSE is not defined, a default value of 3 is
assumed.
.SH ESPS HEADERS
.PP
If the \fB\-w\fP option is specified,
values in the header of 
.I outfile.sd
are copied from  the values in the header of 
.I infile.sd.
In addition,
the generic header item \fIstart_time\fP (type DOUBLE)
is written in the output file.  
The value written is computed by taking the \fIstart_time\fP value from
the header of the input file (or zero, if such a header item doesn't
exist) and adding to it the offset time (from the
beginning of the input file) of the starting point of the word boundry.
.SH "FUTURE CHANGES"
.PP
Have parameters read from a parameter file.
.SH WARNINGS
.PP
\fIFind_ep\fR warns and exits if the input file is not a FEA_SD 
(or old style SD) file, 
if the starting point is not in the file, or if the end of file is 
encountered.
.SH EXAMPLE
.PP
This example uses the defaults and writes out a file \fIutterance.sd\fP
which contains the utterance extracted from the file \fIfile.sd\fP.
File \fIfile.sd\fP contains the utterance and surrounding silence,
while \fIutterance.sd\fP contains only the utterance. The flag -b 
signifies that the file was digitized with 16-bit resolution.  
.nf

         %find_ep -b 16 -w file.sd utterance.sd

.fi
.SH "SEE ALSO"
.PP
copysd(1\-ESPS)
.SH "BUGS"
.PP
None known.
.SH REFERENCES
.PP
[1] L. Rabiner and M. Sambur, "An Algorithm for Determining the Endpoints
of Isolated Utterances," \fIBell. Syst. Tech. J.\fR, Vol. 54, pp. 297-315,
Feb. 1975
.PP
[2] L. Lamel \fIet. al.\fR, "An Improved Endpoint Detector for Isolated Word
Recognition," \fIIEEE Trans. Acoustics, Speech, and Signal Processing\fR,
vol. ASSP-29, pp. 777-785, Aug. 1981
.SH "AUTHOR"
.PP
Manual page by David Burton; code by David Burton.