models.tex

隐马尔科夫模型工具箱

notation to that used for tied-mixture HMMs.  A discrete HMM\index{discrete HMMs} can
have multiple data streams but the width of each stream must be
1.  The output probabilities are stored as logs in a scaled
\index{discrete HMM!output probability scaling}
integer format such that if $d_{js}[v]$ is the stored  discrete
probability  for symbol $v$ in stream $s$ of state $j$, the true
probability is given by
\hequation{
  P_{js}[v] = exp(-d_{js}[v]/2371.8)
}{dpscale}
Storage in the form of scaled logs allows discrete probability
HMMs to be implemented very efficiently since \HTK\ tools
mostly use log arithmetic  and direct storage in log form
avoids the need for a run-time conversion.  The
range determined by the constant 2371.8 
was selected to enable probabilities from 1.0 down to
0.000001 to be stored.


\putprog{tmixhmm2}{85}{HMM using Repeat Counts}{

\hmmc{h}{htm} \\
\hmkw{BeginHMM} \\
\> \hmkw{NumStates} 4 \\
\> \hmkw{State} 2 \hmkw{NumMixes} 5 \\
\> \>   \hmkw{TMix} mix 0.2 0.1 0.3*2 0.1\\
\>\hmkw{State} 3 \hmkw{NumMixes} 5 \\
\>  \>  \hmkw{TMix} mix 0.4 0.3 0.1*3\\
\>\hmkw{TransP} 4 \\
\> \> ... \\
\hmkw{EndHMM} 
}

As an example, Fig~\href{f:dischmm} shows the definition of a  discrete
HMM called \textsf{dhmm1}.  As can be seen, this has two streams.  The codebook
for stream 1 is size 10 and for stream 2, it is size 2.  For consistency with
the representation used for continuous density HMMs, these sizes are encoded 
in the \hmkw{NumMixes}\index{nummixes@$<$NumMixes$>$} specifier.

\mysect{Input Linear Transforms}{lintran}

When reading feature vectors from files HTK will coerce them to the
\texttt{TARGETKIND} specified in the config file. Often the
\texttt{TARGETKIND} will contain certain qualifiers (specifying for
example delta parameters). In addition to this parameter coercion it
is possible to apply a linear transform before, or after, appending
delta, acceleration and third derivative parameters.

\putprog{lintran}{70}{Input Linear Transform}{
\hmmc{j}{lintran.mat} \\
\hmkw{MMFIdMask} *\\
\hmkw{MFCC} \\
\hmkw{PreQual}\\
\hmkw{LinXform}\\
\> \hmkw{VecSize} 2\\
\> \hmkw{BlockInfo} 1 2\\
\> \hmkw{Block} 1\\
\> \>   \hmkw{Xform} 2 5\\
\> \>   \> 1.0 0.1 0.2 0.1 0.4\\
\> \>   \> 0.2 1.0 0.1 0.1 0.1
}
Figure~\ref{f:lintran}  shows an example linear transform. The
\hmkw{PreQual} keyword specifies that the linear transform
is to be applied before the delta and delta-delta
parameters specified in \texttt{TARGETKIND} are added. The default
mode, no \hmkw{PreQual} keyword, applies the linear transform 
after the addition of the qualifiers.

The linear transform fully supports projection from higher 
number of features to a smaller number of features. In the 
example, the parameterised data must consist of 5 \texttt{MFCC}
parameters\footnote{If C0 or normalised log-energy are added
these will be stripped prior to applying the linear transform}.
The model sets that are generated using this transform have
a vector size of 2.

By default the linear transform is stored with the HMM. This is
achieved by adding the \hmkw{InputXform} keyword and specifying the
transform or macroname. To allow compatibilty with tools only
supporting the old format models it is possible to specify that no
linear transform is to be stored with the model. 
\begin{verbatim}
    # Do not store linear transform
    HMODEL: SAVEINPUTXFORM = FALSE
\end{verbatim}
In addition it is possible to specify the linear transform as a
\htool{HPARM} configuration variable, \texttt{MATRTRANFN}.
\begin{verbatim}
    # Specifying an input linear transform
    HPARM: MATTRANFN = /home/test/lintran.mat
\end{verbatim}
When a linear transform is specified in this form it is not necessary
to have a macroname linked with it. In this case the filename
will be used as the macroname (having stripped the directory name)

\mysect{Tee Models}{teemods}

Normally, the transition probability from the non-emitting entry
state to the non-emitting exit state of a HMM will be zero to ensure
that the HMM aligns with at least one observation vector.  
Models which have a non-zero entry to exit transition probability 
are referred to as {\it tee-models}.

Tee-models\index{tee-models} are useful for modelling optional transient effects
such as short pauses and noise bursts, particularly between words.

Although most \HTK\ tools support tee-models, they are incompatible with
those that work with isolated models such as \htool{HInit} and
\htool{HRest}. When a tee-model is loaded into one of these tools, its
entry to exit transition probability is reset to zero and the first row of
its transition matrix is renormalised.

\putprog{tmixhmm}{80}{Tied-Mixture HMM}{
\hmmc{h}{htm} \\
\hmkw{BeginHMM} \\
\> \hmkw{NumStates} 4 \\
\> \hmkw{State} 2 \hmkw{NumMixes} 5 \\
\> \>   \hmkw{TMix} mix 0.2 0.1 0.3 0.3 0.1\\
\>\hmkw{State} 3 \hmkw{NumMixes} 5 \\
\>  \>  \hmkw{TMix} mix 0.4 0.3 0.1 0.1 0.1\\
\>\hmkw{TransP} 4 \\
\> \>  0.0 1.0 0.0 0.0 \\
\>  \>  0.0 0.5 0.5 0.0 \\
\>  \>  0.0 0.0 0.6 0.4 \\
\> \>   0.0 0.0 0.0 0.0 \\
\hmkw{EndHMM} 
}

\mysect{Regression Class Trees for Adaptation}{regtreemods}

In order to perform adaptation \HTK\ generally \index{adaptation!regression tree}requires the use of a
binary regression tree. Its use in the adaptation process is explained
in further detail in chapter~\ref{c:Adapt}.
After construction (see
section~\ref{s:hhedregtree}) the terminal nodes of the binary
regression tree contain mixture component groupings or
clusters. These clusters are referred to as regression base
classes. Each mixture component in an HMM set belongs to a unique
regression base class. The binary regression tree is stored as part of
the HMM set, since its structure is necessary for the dynamic
adaptation procedure described in section~\ref{s:reg_classes}. 
Also each mixture component has a regression base class identifier (the
terminal node indices) stored with it. 

An example is shown in
figure~\ref{f:macregtreedef} and corresponds with the tree shown in
figure~\ref{f:regtree1}. The example shows the use of the keyword {\sf
$<$HMMSetId$>$} used to store an identifier for this HMM set. This is
important because the regression tree is built based on this HMM set
and is hence specific to it. Many sets of transforms may be built that
can be applied to this HMM set, but only one HMM set can be
transformed by a transform set that utilises the regression tree.
The regression tree is described by non-terminal nodes {\sf $<$Nodes$>$} and
terminal nodes {\sf $<$TNodes$>$}. Each node contains its index
followed by either the indices of its children (if it is a
non-terminal) or the number of mixture components clustered at a
terminal. Each mixture component as defined by the keyword  {\sf
$<$Mixture$>$} has an {\sf $<$RClass$>$} keyword followed by the
regression base class index.
When an HMM definition is loaded, a check is made to see that all the
regression classes have been defined and that the total number of
mixture components loaded for each regression class matches the number
of mixture components defined in the regression tree.

The regression tree together with the mixture regression base class
numbers can be constructed automatically with the use of the tool
\htool{HHEd} (see section~\ref{s:hhedregtree}).


\mysect{Binary Storage Format}{binsave}

Throughout this chapter, a text-based representation
has been used for the external storage of HMM
definitions.  For experimental work, text-based storage allows
simple and direct 
access to HMM parameters and this can be invaluable.
However, when using very large HMM sets, storage in text form
is less practical since it is inefficient in its use of
memory and the time taken to load can be excessive due to
the large number of character to float conversions needed.

To solve these problems, \HTK\ also provides a 
binary storage\index{HMM definition!binary storage}\index{binary storage}
format.  In binary mode, keywords are written as a single
colon followed by an 8 bit code representing the actual
keyword.  Any subsequent numerical information following
the keyword is then in binary.  Integers are written as
16-bit shorts and all floating-point numbers are written
as 32-bit single precision floats.  The repeat
factor used in the run-length encoding
scheme for tied-mixture and discrete HMMs is written as
a single byte.  Its presence immediately after a 16-bit
discrete log probability is indicated by setting the top
bit to 1 (this is the reason why the range of discrete 
log probabilities is limited to
0 to 32767 i.e.\ only 15 bits are used for the actual
value).  For tied-mixtures, the repeat count is signalled
by subtracting 2.0 from the weight.

Binary storage format and text storage format can be mixed
within and between input files.  Each time a keyword is
encountered, its coding is used to determine whether the
subsequent numerical information should be input in text
or binary form.  This means, for example, that binary
files can be manually patched by replacing a binary-format
definition by a text format definition\footnote{The fact that
this is possible does not mean that it is recommended practice!}.

\HTK\ tools provide a standard command line option (\texttt{-B}) to indicate
that HMM definitions should be output in binary format.
Alternatively, the Boolean configuration 
variable \texttt{SAVEBINARY}\index{savebinary@\texttt{SAVEBINARY}} can be set to true to
force binary format output.

\putprog{dischmm}{80}{Discrete Probability HMM}{
\hmmt{o} \hmkw{DISCRETE} \hmkw{StreamInfo} 2 1 1 \\
\hmmc{h}{dhmm1} \\
\hmkw{BeginHMM} \\
\> \hmkw{NumStates} 4  \\
\> \hmkw{State} 2   \\
\>\>   \hmkw{NumMixes} 10 2 \\
\>\>   \hmkw{SWeights} 2 0.9 1.1 \\
\>\>   \hmkw{Stream} 1 \\
\> \>\>     \hmkw{DProb} 3288*4  32767*6 \\
\>\>   \hmkw{Stream} 2 \\
\>\> \>     \hmkw{DProb} 1644*2 \\
\> \hmkw{State} 3   \\
\>\>   \hmkw{NumMixes} 10 2 \\
\>\>   \hmkw{SWeights} 2 0.9 1.1 \\
\> \>  \hmkw{Stream} 1 \\
\> \> \>    \hmkw{DProb} 5461*10 \\
\> \>  \hmkw{Stream} 2 \\
\> \> \>    \hmkw{DProb} 1644*2 \\
\>\hmkw{TransP} 4 \\
\> \>  0.0 1.0 0.0 0.0 \\
 \> \>  0.0 0.5 0.5 0.0 \\
 \> \>  0.0 0.0 0.6 0.4 \\
\> \>   0.0 0.0 0.0 0.0 \\
\hmkw{EndHMM} 
}

\putprog{macregtreedef}{80}{MMF with a regression tree and classes}
{
\hmmt{o} \> \hmkw{HMMSetId} ecrl\_us\_mono \\
\>    \hmkw{VecSize} 4 \hmkw{MFCC} \\
\hmmc{r}{ecrl\_us\_mono\_tree\_4} \\
\>    \hmkw{RegTree} 4 \\
\>    \hmkw{Node} 1 2 3 \\
\>    \hmkw{Node} 2 4 5 \\
\>    \hmkw{Node} 3 6 7 \\
\>    \hmkw{TNode} 4 30 \\
\>    \hmkw{TNode} 5 25 \\
\>    \hmkw{TNode} 6 40 \\
\>    \hmkw{TNode} 7 39 \\
\hmmc{s}{stateA} \hmkw{NumMixes} 3 \\
\>    \hmkw{Mixture} 1 0.34 \\
\>\>    \hmkw{RClass} 4 \\
\>\>    \hmmc{u}{mean51} \\   
\>\>    \hmmc{v}{var65} \\
\>    \hmkw{Mixture} 2 0.52 \\
\>\>    \hmkw{RClass} 7 \\
\>\>    \hmmc{u}{mean32} \\   
\>\>    \hmmc{v}{var65} \\
\>    \hmkw{Mixture} 3 0.14 \\
\>\>    \hmkw{RClass} 5 \\
\>\>    \hmmc{u}{mean12} \\   
\>\>    \hmmc{v}{var3}
}

\mysect{The HMM Definition Language}{hmmdef}

To conclude this chapter,
this section presents a formal\index{HMM definition!formal syntax} description
of the HMM definition language used by \HTK.
Syntax is described using an extended BNF notation in which
alternatives are separated by a vertical bar $|$, parentheses () denote
factoring, brackets [\ ] denote options, and braces \{\} denote zero or more
repetitions. 

All keywords are enclosed in angle brackets\footnote{
This definition covers the textual version only.  The syntax for
the binary format
is identical apart from the way that the lexical items are encoded.} and
the case of the
keyword name is not significant.
White space is not significant except within double-quoted strings.

The top level structure of a HMM definition is shown by the following
rule. 
{\sf
\begin{tabbing}
++++ \= ++++++++ \= ++ \= +++++++++++++++++ \= +++ \=  \kill
\>  hmmdef = \> [ $\sim$h macro ] \\
\>\>         $<$BeginHMM$>$ \\
\>\>\>          [ globalOpts ] \\
\>\>\>          $<$NumStates$>$ short \\
\>\>\>          state \{ state \} \\
\>\>\>          [ regTree ] \\
\>\>\>          transP \\
\>\>\>          [ duration ] \\
\>\>         $<$EndHMM$>$ 
\end{tabbing}
}
A HMM definition consists of an optional set of global options\index{HMM definition!global options} followed by
the \hmkw{NumStates}\index{numstates@$<$NumStates$>$} keyword whose following argument specifies the number of states in 
the model inclusive of the non-emitting entry and exit states\footnote{
Integer numbers are specified as either \textsf{char} or \textsf{short}.
This has no effect on text-based definitions but for binary format it indicates
the underlying C type used to represent the number.}.
The information for each state is then given in turn, followed by the 
parameters of the transition matrix and the model duration parameters, if any.
The name of the HMM is given by the \hmmt{h} macro.  If the HMM is the
only definition within a file, the \hmmt{h} macro name can be omitted
and the HMM name is assumed to be the same as the file name.

The global options\index{global options} are common to all HMMs.  They can be given 
separately using a \hmmt{o} option macro