openviron.tex

该压缩包为最新版htk的源代码,htk是现在比较流行的语音处理软件,请有兴趣的朋友下载使用

all of the remaining parameter definitions can be seen to be masked by
higher level over-rides.

The configuration file itself consists of a sequence of parameter
definitions of the form\index{configuration files!format}
\begin{verbatim}
    [MODULE:] PARAMETER = VALUE
\end{verbatim}
One parameter definition is written per line and square brackets
indicate that the module name is
optional. Parameter definitions are not case sensitive
but by convention they are written in upper case.  A \verb+#+ character 
indicates that the rest of the line is a comment.

As an example, the following is a simple configuration file
\begin{verbatim}
    # Example config file
            TARGETKIND = MFCC
            NUMCHANS   = 20
            WINDOWSIZE = 250000.0    # ie 25 msecs
            PREEMCOEF = 0.97
            ENORMALISE = T
    HSHELL: TRACE = 02               # octal
    HPARM:  TRACE = 0101
\end{verbatim}
The first five lines contain no module name and hence they apply
globally, that is, any library module or tool which is interested
in the configuration parameter \texttt{NUMCHANS} will read the given
parameter value.  In practice, this is not a problem with library modules
since nearly all configuration parameters have unique
names.  The final two lines show the same parameter name being given
different values within different modules.  This is an example of
a parameter which every module responds to and hence does not have a unique
name.

This example also shows each of the four possible types of value that can
appear in a configuration file: string\index{string values}, 
integer\index{integer values}, float\index{float values} and 
Boolean\index{Boolean values}.
The configuration parameter \texttt{TARGETKIND}
\index{targetkind@\texttt{TARGETKIND}} requires a string
value specifying the name of a speech parameter kind. Strings not
starting with a letter should be enclosed in double quotes. 
\texttt{NUMCHANS}\index{numchans@\texttt{NUMCHANS}}  requires an 
integer value specifying the number of
filter-bank channels to use in the analysis.
\texttt{WINDOWSIZE}\index{windowsize@\texttt{WINDOWSIZE}}  actually requires a floating-point value
specifying  the window size in units of 100ns.  However, an integer
can always be given wherever a float is required.
\texttt{PREEMCOEF} also requires a floating-point  value specifying the
pre-emphasis coefficient to be used.  Finally, \texttt{ENORMALISE}
\index{enormalise@\texttt{ENORMALISE}} is
a Boolean parameter  which determines whether or not energy
normalisation is to be performed, its value must be \texttt{T}, \texttt{TRUE} or
\texttt{F}, \texttt{FALSE}.  Notice also that, as in command line options,
integer values can use the C conventions for writing in non-decimal bases.
Thus, the trace value of 0101 is equal to decimal 65.  This is particularly
useful in this
case because trace values are typically interpreted as bit-strings by
\HTK\ modules and tools.\index{configuration files!types}

If the name of a configuration variable is mis-typed, there will be no
warning and the variable will simply be ignored.  To help guard
against this, the standard option \texttt{-D} can be used.  This
displays all of the configuration variables before and after the tool
runs.  In the latter case, all configuration variables which are still
unread are marked by a hash character.  The initial display allows the
configuration values to be checked before potentially wasting a large
amount of cpu time through incorrectly set parameters.  The final
display shows which configuration variables were actually used during
the execution of the tool.  The form of the output is shown by the
following example
\begin{verbatim}
    HTK Configuration Parameters[3]
      Module/Tool     Parameter                  Value
    #                 SAVEBINARY                  TRUE
      HPARM           TARGETRATE         256000.000000
                      TARGETKIND                MFCC_0
\end{verbatim}
Here three configuration parameters have been set but the hash
(\verb+#+) indicates that \texttt{SAVEBINARY} has not been used.
\index{configuration variables!display}
\index{configuration files|)}

\mysect{Standard Options}{stdopts}

\index{standard options} As noted in section~\ref{s:cmdline}, options
consisting of a capital letter are common across all tools.  Many are
specific to particular file types and they will be introduced as they
arise.  However, there are six options that are standard across all
tools.  Three of these have been mentioned already.  The
option \texttt{-C}\index{standard options!aaac@\texttt{-C}} is used to
specify a configuration file name and the option
\texttt{-S}\index{standard options!aaas@\texttt{-S}} is used to
specify a script file name, whilst the option
\texttt{-D}\index{standard options!aaad@\texttt{-D}} is used to
display configuration settings.

The two remaining standard options provided directly by 
\htool{HShell} are \texttt{-A}\index{standard options!aaaa@\texttt{-A}}
and \texttt{-V}. 
The option \texttt{-A} causes the current command line arguments to 
be printed. When running experiments via
scripts, it is a good idea to use this option to record in a log file the
precise settings used for each tool.
The option \texttt{-V}\index{standard options!aaav@\texttt{-V}}
causes version information for the tool and each module used by that
tool to be listed.  These should always be quoted when making bug reports.

Finally, all tools implement the trace 
option \texttt{-T}\index{standard options!aaat@\texttt{-T}}.
Trace values are typically bit strings and the meaning of each bit
is described in the reference section for each tool.  Setting a trace\index{tracing}
option via the command line overrides any setting for that same trace
option in a configuration file.  This is a general rule, command line
options always override defaults set in configuration files.

All of the standard options are listed in the final summary section of
this chapter. As a general rule, you should consider passing at least
-A -D -V -T 1 to all tools, which will guarantee that sufficient
information is available in the tool output.


\mysect{Error Reporting}{erep}

The \htool{HShell} module provides a standard mechanism for reporting
errors\index{errors} and warnings\index{warnings}.  
A typical error message is as follows
\begin{verbatim}
    HList: ERROR [+1110]
      IsWave: cannot open file speech.dat
\end{verbatim}
This indicates that the tool \htool{HList} is reporting an error
number +1110.  All errors have positive error 
numbers\index{error numbers!structure of} and always
result in the tool terminating.  Warnings have negative error numbers
and the tool does not terminate.  The first two digits of an error
number indicate the module or tool in which the error is located
(\htool{HList} in this case)
and the last two digits define the class of error.
The second line of the error message names the actual routine
in which the error occurred (here \texttt{IsWave}) and the 
actual error message.    All errors and warnings are listed
in the reference section at the end of this book indexed by
error/warning number.  This listing contains more details on each
error or warning along with suggested causes.

Error messages are sent to the  standard error stream but warnings
are sent to the standard output stream.  The reason for the latter
is that most 
\HTK\ tools are run with progress tracing enabled.  Sending warnings
to the standard output stream ensures that they are properly
interleaved with the trace of progress so that it is easy to determine
the point at which the warning was issued.  Sending warnings to
standard error would lose this information.

The default behaviour of a \HTK\ tool on terminating due to an
error is to exit normally returning the error number as exit status.
If, however, the configuration variable 
\texttt{ABORTONERR}\index{abortonerr@\texttt{ABORTONERR}} is set to
true then the tool will core dump.  This is a debugging facility which
should not concern most users.\index{termination}

\mysect{Strings and Names}{htkstrings}

Many \HTK\  definition files include names of various types of
objects: for example labels, model names, words, etc.
In order to achieve some uniformity, \HTK\ applies standard
rules for reading strings which are names.\index{strings!rules for}
These rules are not, however, necessary when using the language
modelling tools -- see below.

A name string consists of a single white space delimited word or
a quoted string.  Either the single quote \verb+'+ or 
the double quote \verb+"+ can be used to quote strings but the
start and end quotes must be matched.  The backslash \verb+\+ 
character can also
be used to introduce otherwise reserved characters.  The 
character following a backslash is inserted into the string without special
processing unless that character is a digit in the range 0 to 7.  
In that case, the three
characters following the backslash are read and interpreted as an octal
character code.  When the three characters are not octal digits the result
is not well defined.\index{strings!metacharacters in}

In summary the special processing is
\begin{center}
\begin{tabular}{|c|l|} \hline
Notation   & Meaning \\ \hline
\verb+\\+   & \verb+\+ \\ \hline
\verb+\_ +  & represents a space that will not terminate a string \\ \hline
\verb+\'+   & \verb+'+  (and will not end a quoted string) \\ \hline
\verb+\"+   & \verb+"+ (and will not end a quoted string) \\ \hline
\verb+\nnn+ & the character with octal code \verb+\nnn+ \\ \hline
\end{tabular}
\end{center}

\noindent\index{non-printing chars}
Note that the above allows the same effect to be achieved in a number of 
different ways.  For example,
\begin{verbatim}
    "\"QUOTE" 
    \"QUOTE 
    '"QUOTE' 
    \042QUOTE 
\end{verbatim}
all produce the string \verb+"QUOTE+.

The only exceptions to the above general rules are:
\begin{itemize}

\item  Where models are specified in \htool{HHEd} scripts,
commas (\verb+,+), 
dots (\verb+.+),
and closing brackets (\verb+)+) 
are all used as extra delimiters to allow  \htool{HHEd} scripts
created for earlier versions of \HTK\ to be used unchanged.
Hence for example, \texttt{(a,b,c,d)} would be split into 4 
distinct name strings \texttt{a}, \texttt{b}, \texttt{c} and \texttt{d}.

\item When the configuration variable
\texttt{RAWMITFORMAT} is set true,  each word in a language model
definition file consists of a white space delimited string with no 
special processing being performed.

\item  Source dictionaries read by \htool{HDMan} are read using
the standard \HTK\ string conventions, however, the command \texttt{IR}
can be used in a \htool{HDMan} source edit script to switch to using
this raw format.

\item 
To ensure that the general definition of a name string works
properly in \HTK\ master label files, all
MLFs must have the reserved \texttt{.} and \verb+///+ terminators 
alone on a line with no surrounding white space. 
If this causes problems reading old MLF files, the configuration
variable \texttt{V1COMPAT} should be set true in the module \htool{HLabel}.  
In this case,
\HTK\ will attempt to simulate the behaviour of the older version 1.5.

\item
To force numbers to be interpreted as strings rather than times or scores in a
label file, they must be quoted.  If the configuration variable
\texttt{QUOTECHAR} is set to \verb+'+ or \verb+"+ then output labels will be
quoted with the specified quote character.  If \texttt{QUOTECHAR} is set to  
\verb+\+, then output labels will be escaped. The default is to select the 
simplest quoting mechanism.\index{strings!output of}
\end{itemize}

Note that under some versions of \texttt{Unix} \HTK\ can support the 8-bit
character sets used for the representation of various orthographies. In
such cases the shell environment variable \texttt{\$LANG} usually governs
which ISO character set is in use.

\subsubsection{Language modelling tools}