hhed.tex

隐马尔可夫模型源代码

%/* ----------------------------------------------------------- */
%/*                                                             */
%/*                          ___                                */
%/*                       |_| | |_/   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 and Julian Odell  20/05/97
%
\newpage
\mysect{HHEd}{HHEd}

\subsection{Function}

\index{hhed@\htool{HHEd}|(}
\htool{HHEd} is a script driven editor for manipulating sets of HMM definitions.
Its basic operation is to load in a set of HMMs, apply a sequence of edit
operations and then output the transformed set. \htool{HHEd} is mainly used for
applying tyings across selected HMM parameters.  It also has facilities for
cloning HMMs, clustering states and editing HMM structures.

Many \htool{HHEd} commands operate on sets of similar items selected from the
set of currently loaded HMMs.  For example, it is possible to define a set of
all final states of all vowel models, or all mean vectors of all mixture
components within the model X, etc.  Sets such as these are defined by item
lists using the syntax rules given below.  In all commands, all of the items in
the set defined by an item list must be of the same type where the possible
types are
\begin{tabbing}
+++++++ \= ++ \=  ++++++++++++++++ \= ++ \= \kill
\>  s \> -- state             \> t \> -- transition matrix \\
\>  p \> -- pdf               \> w \> -- stream weights \\
\>  m \> -- mixture component \> d \> -- duration parameters \\
\>  u \> -- mean vector       \> x \> -- transform matrix \\
\>  v \> -- variance vector   \> i \> -- inverse covariance matrix \\
\>  h \> -- HMM definition
\end{tabbing}
\noindent
Most of the above correspond directly to the tie points shown in
Fig~\href{f:hierarch}.  There is just one exception.  The type ``p''
corresponds to a pdf (ie a sum of Gaussian mixtures).  Pdf's cannot
be tied, however, they can be named in a Tie (\texttt{TI}) command
(see below) in which case, the effect is to join all of the contained
mixture components into one pool of mixtures and then all of the
mixtures in the pool are shared across all pdf's.  This allows
conventional {\it tied-mixture} or {\it semi-continuous} HMM systems
to be constructed.

The syntax rules for item lists are as follows.  An item list
consists of a comma separated list of item sets.
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>     itemList\>  = \> ``\{'' itemSet \{ ``,'' itemSet \} ``\}'' 
\end{tabbing}}
\noindent
Each {\sf itemSet} consists of the name of one
or more HMMs (or a pattern
representing a set of HMMs) followed by a specification
which represents a set of paths down the parameter hierarchy 
each terminating at one of the required parameter items.
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>      itemSet \>  = \> hmmName . [``transP'' $|$ ``state'' state ]\\
\>      hmmName \>  = \> ident $|$ identList \\
\>     identList\>  = \> ``('' ident \{ ``,'' ident \} ``)'' \\
\>     ident    \>  = \> $<$ char $|$ metachar $>$ \\
\>     metachar \>  = \> ``?'' $|$ ``$\star$''
\end{tabbing}}
\noindent
A {\sf hmmName} consists of a single {\sf ident} or
a comma separated list of {\sf ident}'s.  The following examples
are all valid {\sf hmmName}'s:
\begin{verbatim}
         aa  three  model001  (aa,iy,ah,uh)  (one,two,three)
\end{verbatim}
In addition, an {\sf ident} can contain the metacharacter
``?'' which matches any single character and the metacharacter
``$\star$'' which matches a string of zero or more characters.
For example, the item list
\begin{verbatim}
         {*-aa+*.transP}
\end{verbatim}
would represent the set of transition matrices of all loaded
triphone variations of \texttt{aa}.

Items within states require the state indices to be specified
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>      state \>  = \> index [``.'' stateComp ] \\
\>      index \>  = \> ``['' intRange \{ ``,'' intRange \} ``]'' \\
\>    intRange \> = \> integer [ ``-'' integer ]
\end{tabbing}}
For example, the item list
\begin{verbatim}
         {*.state[1,3-5,9]}
\end{verbatim}
represents the set of all states 1, 3 to 5 inclusive and 9 of all
currently loaded HMMs.  Items within states include durational
parameters, stream weights, pdf's and all items within mixtures
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>   stateComp \>  = \> ``dur'' $|$ ``weights'' $|$ [ `` stream'' index ]
        ``.'' ``mix'' [ mix ]
\end{tabbing}}
For example, 
\begin{verbatim}
         {(aa,ah,ax).state[2].dur}
\end{verbatim}
denotes the set of durational parameter vectors from state 2 of the
HMMs \texttt{aa}, \texttt{ah} and \texttt{ax}.  Similarly, 
\begin{verbatim}
         {*.state[2-4].weights}
\end{verbatim}
denotes the set of stream weights for states 2 to 4 of all currently
loaded HMMs.  The specification of pdf's may optionally include a
list of the relevant streams, if omitted, stream 1 is assumed.
For example,
\begin{verbatim}
         {three.state[3].mix}
\end{verbatim}
and
\begin{verbatim}
         {three.state[3].stream[1].mix}
\end{verbatim}
both denote a list of the single pdf belonging to stream 1 of state 3 
of the HMM \texttt{three}.

Within a pdf, the possible item types are mixture components,
mean vectors, and the various possible forms of covariance
parameters
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>   mix \>  = \>  index [ ``.'' ( ``mean'' $|$ ``cov'' ) ]
\end{tabbing}}
For example,
\begin{verbatim}
         {*.state[2].mix[1-3]}
\end{verbatim}
denotes the set of mixture components 1 to 3 from state 2 of all
currently loaded HMMs and
\begin{verbatim}
         {(one,two).state[4].stream[3].mix[1].mean}
\end{verbatim}
denotes the set of mean vectors from mixture component 1, stream 3,
state 4 of the HMMs \texttt{one} and \texttt{two}.  When {\sf cov}
is specified, the type of the covariance item referred to is 
determined from the \texttt{CovKind} of the loaded models.  Thus,
for diagonal covariance models, the item list
\begin{verbatim}
         {*.state[2-4].mix[1].cov}
\end{verbatim}
would denote the set of variance vectors for mixture 1, states 2 to 4
of all loaded HMMs.

Note finally, that it is not an error to specify non-existent models,
states, mixtures, etc.  All item list specifications are regarded
as patterns which are matched against the currently loaded set of
models.  All and only those items which match are included in
the set.  However, both a null result and a set of items of mixed
type do result in errors.

All \htool{HHEd} commands consist of a 2 character command name followed
by zero or more arguments.  In the following descriptions, item
lists are shown as \texttt{itemList(c)} where the character c denotes the
type of item expected by that command.  If this type indicator is
missing then the command works for all item types.

The \htool{HHEd} commands are as follows

\subsubsection{\texttt{AT i j prob itemList(t)}}

Add a transition from state \texttt{i} to state \texttt{j} with probability
\texttt{prob} for all transition matrices in \texttt{itemList}.  The remaining
transitions out of state \texttt{i} are rescaled so that $\sum_k a_{ik} = 1 $.
For example,
\begin{verbatim}
         AT 1 3 0.1 {*.transP}
\end{verbatim}
would add a skip transition to all loaded models from state 1 to state 3
with probability 0.1.

\subsubsection*{\tt AU hmmList}

Use a set of decision trees to create a new set of models specified
by the \texttt{hmmList}.  The decision trees may be made as a result
of either the \texttt{TB} or \texttt{LT} command.  

Each model in \texttt{hmmList} is constructed in the following manner.
If a model with the same logical name already exists in the
current  HMM set this is used unchanged, otherwise the model is
synthesised from the decision trees.  If the trees cluster at the 
model level the synthesis results in a logical model sharing the
physical model from the tree that matches the new context.  If
the clustering was performed at the state level a prototype model
(an example of the same phone model occurring in a different context)
is found and a new HMM is constructed  that shares the transition
matrix with the prototype model but consists of tied states selected
using the decision tree.

\subsubsection*{\tt CL hmmList}

Clone a HMM list.  The file \texttt{hmmList} should hold a list of HMMs
all of whose logical names are either the same as, or are context-dependent
versions of the currently loaded set of HMMs.  For each name in \texttt{hmmList},
the corresponding HMM in the loaded set is cloned.  On completion, the
currently loaded set is discarded and replaced by the new set.  For example,
if the file \texttt{mylist} contained
\begin{verbatim}
         A-A+A
         A-A+B
         B-A+A
         B-B+B
         B-B+A
\end{verbatim}
and the currently loaded HMMs were just \texttt{A} and \texttt{B}, then
\texttt{A} would be cloned 3 times to give the models \texttt{A-A+A},
\texttt{A-A+B} and \texttt{B-A+A}, and \texttt{B} would be cloned 2 times
to give \texttt{B-B+B} and \texttt{B-B+A}.  On completion, the original
definitions for \texttt{A} and \texttt{B} would be deleted (they could
be retained by including them in the new \texttt{hmmList}).

\subsubsection*{\tt CO newList}

Compact a set of HMMs.  The effect of this command is to
scan the currently loaded set of HMMs and identify all identical
definitions.  The physical name of the first model in each identical
set is then assigned to all models in that set and all
model definitions are replaced by a pointer to the first
model definition.  On completion,
a new list of HMMs which includes the new model tyings is
written out to file \texttt{newList}.  For example, suppose that
models \texttt{A}, \texttt{B}, \texttt{C} and \texttt{D} were currently
loaded and \texttt{A} and \texttt{B} were identical.  Then the command
\begin{verbatim}
         CO tlist
\end{verbatim}
would tie HMMs \texttt{A} and \texttt{B}, 
set the physical name of \texttt{B} to \texttt{A} and output the
new HMM list
\begin{verbatim}
         A
         B A
         C
         D
\end{verbatim}
to the file \texttt{tlist}.  This command is used mainly after 
performing a sequence of parameter tying commands.

\subsubsection*{\tt DP s n id ...}

Duplicates a set of HMMs.  This command is used to replicate a
set of HMMs whilst allowing control over which structures will
be shared between them.