labels.tex

隐马尔科夫模型工具箱

%/* ----------------------------------------------------------- */
%/*                                                             */
%/*                          ___                                */
%/*                       |_| | |_/   SPEECH                    */
%/*                       | | | | \   RECOGNITION               */
%/*                       =========   SOFTWARE                  */ 
%/*                                                             */
%/*                                                             */
%/* ----------------------------------------------------------- */
%/*         Copyright: Microsoft Corporation                    */
%/*          1995-2000 Redmond, Washington USA                  */
%/*                    http://www.microsoft.com                */
%/*                                                             */
%/*   Use of this software is governed by a License Agreement   */
%/*    ** See the file License for the Conditions of Use  **    */
%/*    **     This banner notice must not be removed      **    */
%/*                                                             */
%/* ----------------------------------------------------------- */
%
% HTKBook - Steve Young 1/12/97
%

\mychap{Transcriptions and Label Files}{labels}

\sidepic{Tool.labs}{80}{}
Many of the operations performed by \HTK\ which involve speech data
files assume that the speech is divided into segments and each segment
has a name or \textit{label}.  The set of labels associated with a
speech file constitute a \textit{transcription} and each transcription is
stored in a separate \textit{label file}.  Typically, the name of the
label file will be the same as the corresponding speech file but with
a different extension.  For convenience, label files are often stored
in a separate directory and all \HTK\ tools have an option to specify
this.  When very large numbers of files are being processing, label
file access can be greatly facilitated by using\index{master label files}
\textit{Master Label Files (MLFs)}.  MLFs may be regarded as index\index{MLF}
files holding pointers to the actual label files which can either be
embedded in the same index file or stored anywhere else in the file system.
Thus, MLFs allow large sets of files to be stored in a single file, they
allow a single transcription to be shared by many logical label files
and they allow arbitrary file redirection.\index{label files}

The \HTK\ interface to label files is provided by the module \htool{HLabel}
which implements the MLF facility and support for a number of external
label file formats. 
All of the facilities supplied by \htool{HLabel}, including the
supported label file formats, are described in this chapter.
In addition, \HTK\ provides a tool called \htool{HLEd} for simple 
batch editing of label files and this is also described.
Before proceeding to the details, however, the general structure of
label files will be reviewed.

\mysect{Label File Structure}{labstruct}

Most transcriptions are single-alternative and single-level, that is
to say, the associated speech file is described by a single sequence
of labelled segments.  Most standard label formats are of this kind.
Sometimes, however, it is useful to have several levels of labels associated
with the same basic segment sequence.   For example, in training a HMM
system it is useful to have both the word level transcriptions and the
phone level transcriptions \textit{side-by-side}.  \index{labels!side-by-side}

Orthogonal to the requirement for multiple levels of description,
a transcription may also need to include multiple alternative
descriptions of the same speech file.  For example, the output
of a speech recogniser may be in the form of an \textit{N-best} list
where each word sequence in the list represents one possible interpretation
of the input.\index{labels!multiple level}

As an example, Fig.~\href{f:labegs} shows a speech file and three
different ways in which it might be labelled.  In part (a), just a simple
orthography is given and this single-level single-alternative type of
transcription is the  commonest case. Part (b) shows a 2-level
transcription  where the basic level consists of a sequence of phones but
a higher level of word labels are also provided.  Notice that there is a
distinction between the basic level and the higher levels, since only the
basic level has explicit boundary locations marked for every segment. 
The higher levels do not have explicit boundary information since this
can always be inferred from the basic level boundaries. Finally, part (c)
shows the case where knowledge of the contents of the speech file is
uncertain and three possible word sequences are given.

\HTK\ label files support multiple-alternative and multiple-level
transcriptions.  In addition to start and end times on the basic level, a
label at  any level may also have a score associated with it.  When a
transcription is loaded, all but one specific alternative can be discarded by
setting the configuration variable \texttt{TRANSALT}\index{transalt@\texttt{TRANSALT}} to the required
alternative \texttt{N}, where the first (i.e. normal) alternative is numbered
1.  Similarly, all but a specified level can be discarded by setting the
configuration variable 
\texttt{TRANSLEV}\index{translev@\texttt{TRANSLEV}} to 
the required level number where again the first (i.e.
normal) level is numbered 1.

All non-\HTK\ formats are limited to
single-level single-alternative transcriptions.


\mysect{Label File Formats}{labform}

As with speech data files, \HTK\ not only defines its own format for
label files but also supports a number of external formats.  Defining
an external format is similar to the case for speech data files except
that the relevant configuration variables for specifying a format
other than \HTK\ are called \texttt{SOURCELABEL}\index{sourcelabel@\texttt{SOURCELABEL}} and \texttt{TARGETLABEL}.
The source label format can also be specified using the 
\texttt{-G}\index{standard options!aaag@\texttt{-G}} command
line option.  As with using the 
\texttt{-F}\index{standard options!aaaf@\texttt{-F}} command
line option for speech data files, the \texttt{-G} option overrides any
setting of \texttt{SOURCELABEL}\index{labels!external formats}

\subsection{HTK Label Files}

The \HTK\ label format is text based.  As noted above, a single label
file can contain multiple-alternatives and multiple-levels.

Each line of a \HTK\ label file contains\index{label files!HTK format}
the actual label optionally preceded by start and end times, and
optionally followed by a match score.  
\begin{verbatim}
    [start  [end] ] name [score] { auxname [auxscore] } [comment]
\end{verbatim}
where \texttt{start} denotes the start time of the labelled segment
in 100ns units, \texttt{end}
denotes the end time in 100ns units,  \texttt{name} is the name
of the segment and \texttt{score} is a floating point confidence score.
All fields except the name are optional.  If \texttt{end} is omitted then
it is set equal to -1 and ignored.  This case would occur with data which had
been labelled frame synchronously.  If \texttt{start} and \texttt{end} are both
missing then both are set to -1 and the label file is treated as a 
simple symbolic transcription.  The
optional score would typically be a log probability generated by a 
recognition tool.  When omitted the score is set to 0.0.

The following example corresponds to the transcription shown
in part (a) of Fig.~\ref{f:labegs}
\begin{verbatim}
    0000000 3600000 ice
    3600000 8200000 cream
\end{verbatim}
Multiple levels are described by adding further names alongside
the basic name.  The lowest level (shortest segments) should be
given first since only the lowest level has start and end times.
The label file corresponding to the transcription illustrated in
part (b) of Fig.~\ref{f:labegs} would be as follows.
\begin{verbatim}
    0000000 2200000 ay     ice
    2200000 3600000 s
    3600000 4300000 k      cream
    4300000 5000000 r
    5000000 7400000 iy
    7400000 8200000 m
\end{verbatim}
Finally, multiple alternatives are written as a sequence of separate
label lists separated by three slashes (///).
The label file corresponding to the transcription illustrated in
part (c) of Fig.~\ref{f:labegs} would therefore be as follows.
\begin{verbatim}
    0000000 2200000 I
    2200000 8200000 scream
    ///
    0000000 3600000 ice
    3600000 8200000 cream
    ///
    0000000 3600000 eyes
    3600000 8200000 cream
\end{verbatim}

Actual label names can be any sequence of characters.
However, the \texttt{-} and \texttt{+} characters are reserved for identifying
the left and right context\index{labels!context markers}, 
respectively, in a context-dependent phone
label.  For example, the label \texttt{N-aa+V} might be used to denote
the phone \texttt{aa} when preceded by a nasal and followed by a vowel.
These context-dependency conventions are used in the label editor \htool{HLEd},
and are understood by all \HTK\ tools.

\subsection{ESPS Label Files}

An \ESPSwaves\  label file is a text file with one label stored per
line. Each label indicates a segment boundary. \index{label files!ESPS format}
A complete description
of the \ESPSwaves\  label format is given in the \ESPSwaves\ 
manual pages {\bf xwaves (1-ESPS)} and {\bf xlabel (1-ESPS)}.
Only details required for use with \HTK\ are given here. 

The label data follows
a header which ends with a line containing only
a \texttt{\#}.  The header contents are generally ignored by \htool{HLabel}.
The labels follow the header in the form
\begin{verbatim}
   time  ccode  name 
\end{verbatim}
where \texttt{time} is a floating point number which denotes the boundary
location  in seconds,
\texttt{ccode} is an integer color map entry used by \ESPSwaves\  in drawing
segment boundaries and \texttt{name} is the name of the segment boundary. A
typical value for \texttt{ccode} is \texttt{121}.

While each \HTK\ label can contain both a start and an end time which
indicate
the boundaries of a labeled segment, \ESPSwaves\ labels
contain a single time in seconds which (by convention) refers 
to the end of the labeled segment.  The starting time
of the segment is taken to be the end of the previous
segment and \texttt{0} initially.

\ESPSwaves\  label files may have several boundary names per line.
However, \htool{HLabel} only reads \ESPSwaves\ label files with a single name
per boundary.  Multiple-alternative and/or multiple-level \HTK\ label
data structures cannot be saved using  \ESPSwaves\ format label files.

\subsection{TIMIT Label Files}

\index{label files!TIMIT format}
TIMIT label files are identical to single-alternative single-level HTK
label files without scores except that the start and end times are
given as sample numbers rather than absolute times.  TIMIT label files
are used on both the prototype and final versions of the TIMIT CD ROM.

\subsection{SCRIBE Label Files}

\index{label files!SCRIBE format}
The SCRIBE label file format is a subset of the European SAM label file format.
SAM label files are text files and each line begins with a label identifying
the type of information stored on that line.  The \HTK\ SCRIBE format recognises
just three label types
\begin{tabbing}
++ \= +++++++ \= \kill
\> LBA \>-- acoustic label \\
\> LBB \>-- broad class label \\
\> UTS \>-- utterance 
\end{tabbing}
For each of these, the rest of the line is divided into comma separated
fields.  The LBA and LBB types have 4 fields: start sample, centre sample, end sample
and label.  \HTK\ expects the centre sample to be blank.  The UTS type has 3 fields:
start sample, end sample and label.  UTS labels may be multi-word since they can
refer to a complete utterance.  In order to make such labels usable within \HTK\ tools,
between word blanks are converted to underscore characters.  The 
\texttt{EX}\index{ex@\texttt{EX} command} command
in the \HTK\ label editor \htool{HLEd} can then be used to split such
a compound label into individual word labels if required.


\mysect{Master Label Files}{mlfs}

\subsection{General Principles of MLFs}

\index{master label files}
Logically, the organisation of data and label files is very simple.
Every data file has a label file of the same name (but
different extension) which is either stored in the same directory as
the data file or in some other specified directory.

\sidefig{labegs}{60}{Example Transcriptions}{2}{}
This scheme is sufficient for most needs and commendably simple.
However, there are many cases where either it makes unnecessarily
inefficient use of the operating system or it seriously inconveniences
the user.  For example, to use a training tool with
isolated word data may require the generation of hundreds or thousands of
label files each having just one label entry.  Even where individual
label files are appropriate (as in the phonetically transcribed TIMIT
database), each label file must be
stored in the same directory as the data file it transcribes, or all label files
must be stored in the same directory.  One cannot, for example, have a
different directory of label files for each TIMIT dialect region and
then run the \HTK\ training tool \htool{HERest} on the whole database.

All of these problems can be solved by the use of Master Label Files
(MLFs).  Every \HTK\ tool which uses label files has a 
\texttt{-I}\index{standard options!aaai@\texttt{-I}} option
which can be used to specify the name of an MLF file.  When an MLF has been
loaded, the normal rules for locating a label file apply except that
the MLF is searched first.  If the required label file \texttt{f} is found via