netdict.tex

Hidden Markov Toolkit (HTK) 3.2.1 HTK is a toolkit for use in research into automatic speech recogn

For example, suppose that decimal number
input was required.  A suitable network structure would be
as shown in Fig.~\href{f:decinet}.  However, to write this directly
in an SLF file would require the digit loop to be written twice.
This can be avoided by defining the digit loop as a sub-network
and referencing it within the main \textit{decimal} network as
follows

\begin{verbatim}
    # Digit network
    SUBLAT=digits
    N=14 L=21
    # define digits
    I=0  W=zero
    I=1  W=one
    I=2  W=two
    ...
    I=9  W=nine
    #  enter/exit & loop-back null nodes
    I=10 W=!NULL
    I=11 W=!NULL
    I=12 W=!NULL
    I=13 W=!NULL
    # null->null->digits
    J=0 S=10 E=11
    J=1 S=11 E=0
    J=2 S=11 E=1
    ...
    J=10 S=11 E=9
    # digits->null->null
    J=11 S=0 E=12
    ...
    J=19 S=9 E=12
    J=20 S=12 E=13
    # finally add loop back
    J=21 S=12 E=11
    .

    # Decimal netork
    N=5 L=4
    # digits -> point -> digits
    I=0 W=start
    I=1 L=digits
    I=2 W=pause
    I=3 L=digits
    I=4 W=end
    # digits -> point -> digits
    J=0 S=0 E=1
    J=1 S=1 E=2
    J=2 S=2 E=3
    J=3 S=3 E=4
\end{verbatim}
The sub-network is identified by the field 
\texttt{SUBLAT}\index{sublat@\texttt{SUBLAT}} in the header
and it is terminated by a single period on a line by itself.  The
main body of the sub-network is written as normal.
Once defined, a sub-network can be substituted into a higher level
network using an \texttt{L} field in a node definition, as in nodes
1 and 3 of the decimal network above.

Of course, this process can be continued and a higher level network
could reference the decimal network wherever it needed decimal
number entry.

\centrefig{bobig}{100}{Back-off Bigram Word-Loop Network}

One of the commonest form of recognition network is the 
word-loop\index{word-loop network}
where all vocabulary items are placed in parallel with a loop-back
to allow any word sequence to be recognised.  This is the basic
arrangement used in most dictation or transcription applications.
\htool{HBuild} can build such a loop automatically from a list
of words.  It can also read in a bigram in either ARPA MIT-LL 
format or HTK matrix format and attach a bigram probability to
each word transition.  Note, however, that using a full bigram
language model means that every distinct pair of words must
have its own unique loop-back transition.  This increases the size of
the network considerably and slows down the recogniser.
When a back-off bigram is used, however, backed-off transitions
can share a common loop-back transition.  Fig.~\href{f:bobig}
illustrates this.  When backed-off bigrams are input via an ARPA MIT-LL 
format file, \htool{HBuild} will exploit this where possible.

Finally, \htool{HBuild} can automatically construct a 
word-pair grammar\index{word-pair grammar} as 
used in the ARPA Naval Resource Management task.


\mysect{Testing a Word Network using \htool{HSGen}}{usehsgen}

When designing task grammars, it is useful to be able to check
that the  language defined by the final word network is as envisaged.
One simple way to check this is to use the network as a generator by
randomly traversing it and outputting the name of each word node
encountered.  \HTK\ provides a very simple tool called 
\htool{HSGen}\index{hsgen@\htool{HSGen}}
for doing this.

As an example if the file \texttt{bnet} contained the simple Bit-But
netword described above and the file \texttt{bdic} contained a corresponding
dictionary then the command
\begin{verbatim}
    HSGen bnet bdic
\end{verbatim}
would generate a random list of examples of the language
defined by  \texttt{bnet}, for example,
\begin{verbatim}
    start bit but bit bit bit end 
    start but bit but but end 
    start bit bit but but end 
    .... etc
\end{verbatim}
This is perhaps not too informative in this case but for more
complex grammars, this type of output can be quite illuminating.

\htool{HSGen} will also estimate the empirical entropy
by recording
the probability of each sentence generated\index{sentence generation}.  
To use this facility, it
is best to suppress the sentence output and generate a large number
of examples.  For example, executing
\begin{verbatim}
    HSGen -s -n 1000 -q bnet bdic
\end{verbatim}
where the \texttt{-s} option requests statistics, the \texttt{-q} option
suppresses the output and \texttt{-n 1000} asks for 1000 sentences
would generate the following output
\begin{verbatim}
    Number of Nodes = 4 [0 null], Vocab Size = 4
    Entropy = 1.156462,  Perplexity = 2.229102
    1000 Sentences: average len = 5.1, min=3, max=19
\end{verbatim}

\mysect{Constructing a Dictionary}{usehdman}

As explained in section~\ref{s:netuse}, the word level network is expanded by
\htool{HNet} to create the network of HMM instances needed by the recogniser.
The way in which each word is expanded is determined from a
dictionary\index{dictionary!construction}.

A dictionary for use in \HTK\ has a very simple format.\index{dictionary!formats}
Each line consists of a single word pronunciation with format
\begin{verbatim}
    WORD [ '['OUTSYM']' ] [PRONPROB] P1 P2 P3 P4 ....
\end{verbatim}
where \texttt{WORD} represents the word, followed by the optional
parameters \texttt{OUTSYM} and \texttt{PRONPROB}, where
\texttt{OUTSYM} is the symbol to output when that word is
recognised (which must be enclosed in square brackets, \verb|[| and
\verb|]|) and \texttt{PRONPROB} is the pronunciation probability
($0.0$ - $1.0$).  \texttt{P1}, \texttt{P2}, \ldots is the sequence of
phones or HMMs to be used in recognising that word. The output symbol
and the pronunciation probability are optional. If an output symbol is
not specified, the name of the word itself is output. If a
pronunciation probability is not specified then a default of 1.0 is
assumed.  Empty square brackets,
\texttt{[]}, can be used to suppress any output when that word is recognised.
For example, a dictionary might contain
\begin{verbatim}
    bit           b  ih t 
    but           b  ah t
    dog    [woof] d  ao g
    cat    [meow] k  ae t
    start  []     sil
    end    []     sil
\end{verbatim}

\noindent
If any word has more than one pronunciation, then the word
has a repeated entry, for example,
\begin{verbatim}
    the           th iy
    the           th ax 
\end{verbatim}
corresponding to the stressed and unstressed forms of the word
``the''.\index{dictionary!output symbols}

The pronunciations in a dictionary are normally at the phone
level as in the above examples.  However, if context-dependent
models are wanted, these can be included directly in the dictionary.
For example, the Bit-But entries might be written as
\begin{verbatim}
    bit           b+ih  b-ih+t  ih-t 
    but           b+ah  b-ah+t  ah-t
\end{verbatim}
In principle, this is never necessary since \htool{HNet} can perform context
expansion automatically, however, it saves computation to do this
off-line as part of the dictionary construction process.  Of course,
this is only possible for word-internal context dependencies.
Cross-word dependencies can only be generated by \htool{HNet}.

\centrefig{dmaker}{110}{Dictionary Construction using \htool{HDMan}}

Pronouncing dictionaries are a valuable resource and if produced
manually, they can require considerable investment.  There are
a number of commercial and public domain dictionaries available,
however, these will typically have differing formats and will
use different phone sets.  To assist in the process of
dictionary construction, \HTK\ provides a tool called \htool{HDMan}
which can be used to edit and merge differing source dictionaries
to form a single uniform dictionary.  The way that
\htool{HDMan}\index{hdman@\htool{HDMan}} works is illustrated in Fig.~\href{f:dmaker}.

Each source dictionary file must have one pronunciation per line and the
words must be sorted into alphabetical order.  The word entries must be
valid \HTK\ strings as defined in section~\ref{s:htkstrings}.  If an
arbitrary character sequence is to be allowed, then the input edit
script should have the command \texttt{IM RAW} as its first command.

The basic operation of 
\htool{HDMan} is to scan the input streams and for each new word
encountered, copy the entry to the output.  In the figure,  a word list
is also shown.  This is optional but if included 
\htool{HDMan} only copies words in the list.  Normally, \htool{HDMan}
copies just the first pronunciation that it finds for any word. Thus,
the source dictionaries are usually arranged in order of
\textit{reliability}, possibly preceded by a small dictionary of special
word pronunciations. For example, in Fig.~\href{f:dmaker}, the main
dictionary might be \texttt{Src2}.  \texttt{Src1} might be a small 
dictionary containing correct pronunciations for words in \texttt{Src2}
known to have  errors in them. Finally, \texttt{Src3} might be a large
poor quality dictionary (for example, it could be generated
by a rule-based text-to-phone system) which is included as a last resort
source of pronunciations for words not in the main dictionary.

As shown in the figure, \htool{HDMan} can apply a set of editing
commands to each source dictionary and it can also edit the
output stream.  The commands available are described in full in
the reference section.  They operate in a similar way to
those in \htool{HLEd}.  Each set of commands is written in
an edit script with one command per line.  Each input edit script
has the same name as the corresponding source dictionary but with
the extension \texttt{.ded} added.  The output edit script is stored
in a file called \texttt{global.ded}\index{global@\texttt{global.ded}}.  
The commands provided
include replace and delete at the word and phone level, context-sensitive
replace and automatic conversions to left biphones, right biphones
and word internal triphones.\index{dictionary!edit commands}

When \htool{HDMan} loads a dictionary it adds word boundary symbols to
the start and end of each pronunciation and then deletes them when
writing out the new dictionary.  The default for these word boundary
symbols is \texttt{\#} but it can be redefined using the \texttt{-b}
option.  The reason for this is to allow context-dependent edit commands 
to take account of word-initial and word-final phone positions.  
The examples below will illustrate this.

Rather than go through each \htool{HDMan} edit command in detail, some examples
will illustrate the typical manipulations that can be performed
by \htool{HDMan}.  Firstly, suppose that a dictionary transcribed
unstressed ``-ed'' endings as \texttt{ih0 d}
but the required dictionary
does not mark stress but uses a schwa in such cases, that is,
the transformations\index{mp@\texttt{MP} command}\index{sp@\texttt{SP} command}
\begin{verbatim}
    ih0 d  #   ->   ax d
    ih0        ->   ih  (otherwise)
\end{verbatim}
are required.
These could be achieved by the following 3 commands
\begin{verbatim}
    MP axd0 ih0 d #
    SP axd0 ax d #
    RP ih ih0
\end{verbatim}
The context sensitive replace is achieved by merging all sequences
of \texttt{ih0 d \#} and then splitting the result into the sequence
\texttt{ax d \#}.  The final \texttt{RP} command\index{rp@\texttt{RP} command} then unconditionally
replaces all occurrences of \texttt{ih0} by \texttt{ih}.
As a second similar example, suppose that all examples of \texttt{ax l}
(as in ``bottle'') are to be replaced by the single phone \texttt{el}
provided that the immediately following phone is a non-vowel.
This requires the use of the \texttt{DC} command\index{dc@\texttt{DC} command} to define a
context consisting of all non-vowels, then a merge using  \texttt{MP}
as above followed by a context-sensitive replace
\begin{verbatim}
    DC nonv l r w y .... m n ng #
    MP axl ax l
    CR el * axl nonv
    SP axl ax l
\end{verbatim}
the final step converts all non-transformed cases of \texttt{ax l}