www.pudn.com > HMM_HTK-3.0.rar > discmods.tex
%/* ----------------------------------------------------------- */
%/* */
%/* ___ */
%/* |_| | |_/ 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 15/11/95
%
\mychap{Discrete and Tied-Mixture Models}{discmods}
\sidepic{Tool.disc}{80}{
Most of the discussion so far has focussed on using \HTK\
to model sequences of continuous-valued vectors. In contrast, this chapter
is mainly concerned with using \HTK\ to model sequences of discrete
symbols.
Discrete symbols arise naturally in modelling many types of
data, for example, letters and words, bitmap images, and DNA sequences.
Continuous signals can also be converted to discrete symbol sequences
by using a quantiser and in particular, speech vectors can be
\textit{vector quantised} as described in section~\ref{s:vquant}.
In all cases, \HTK\ expects a set of $N$ discrete symbols to be represented by
the contiguous sequence of integer numbers from 1 to $N$.
}\index{discrete HMMs}
In \HTK\ discrete probabilities are regarded as being closely analogous
to the mixture weights of a continuous density system. As a consequence,
the representation and processing of discrete HMMs shares a great deal
with continuous density models. It follows from this that most of the
principles and practice developed already are equally applicable to
discrete systems. As a consequence, this chapter can be quite brief.
The first topic covered concerns building HMMs for discrete
symbol sequences. The use of discrete HMMs with speech is then
presented. The tool \htool{HQuant} is described and the method
of converting continuous speech vectors to discrete symbols is
reviewed. This is
followed by a brief discussion of tied-mixture systems which can be
regarded as a compromise between continuous and discrete density systems.
Finally, the use of the \HTK\ tool \htool{HSmooth} for
parameter smoothing by deleted interpolation is presented.
\mysect{Modelling Discrete Sequences}{discseq}
Building HMMs for discrete symbol sequences is essentially the same
as described previously for continuous density systems.
Firstly, a prototype HMM definition must be specified in order
to fix the model topology. For example, the following
is a 3 state ergodic HMM in which the emitting states
are fully connected.
\begin{verbatim}
~o 1 1
~h "dproto"
5
2 10
5461*10
3 10
5461*10
4 10
5461*10
5
0.0 1.0 0.0 0.0 0.0
0.0 0.3 0.3 0.3 0.1
0.0 0.3 0.3 0.3 0.1
0.0 0.3 0.3 0.3 0.1
0.0 0.0 0.0 0.0 0.0
\end{verbatim}
As described in chapter~\ref{c:HMMDefs}, the notation for discrete
HMMs borrows heavily on that used for continuous density models
by equating mixture components with symbol indices. Thus,
this definition assumes that each training data sequence contains
a single stream of symbols indexed from 1 to 10. In this example,
all symbols in each state have been set to be equally likely\footnote{
Remember that discrete probabilities are scaled such that
32767 is equivalent to a probability of 0.000001 and 0 is
equivalent to a probability of 1.0
}. If prior information is available then this can of course be used
to set these initial values.
The training data needed to build a discrete HMM can take one of two forms. It
can either be discrete (\texttt{SOURCEKIND=DISCRETE}) in which case it consists
of a sequence of 2-byte integer symbol indices. Alternatively, it can consist
of continuous parameter vectors with an associated VQ codebook. This latter
case is dealt with in the next section. Here it will be assumed that the data
is symbolic and that it is therefore stored in discrete form.\index{discrete
data} Given a set of training files listed in the script file
\texttt{train.scp}, an initial HMM could be estimated using
\begin{verbatim}
HInit -T 1 -w 1.0 -o dhmm -S train.scp -M hmm0 dproto
\end{verbatim}
This use of \htool{HInit} is identical to that which would be
used for building whole word HMMs where no associated label file is
assumed and the whole of each training sequence is used to estimate
the HMM parameters. Its effect is to read in the prototype
stored in the file \texttt{dproto} and then use the training examples
to estimate initial values for the output distributions
and transition probabilities. This is done by firstly uniformly
segmenting the data and for each segment counting the number of occurrences
of each symbol. These counts are then normalised to provide output distributions
for each state. \htool{HInit} then uses the Viterbi algorithm to resegment
the data and recompute the parameters. This is repeated until convergence
is achieved or an upper limit on the iteration count is reached.
The transition probabilities at each step are estimated simply by
counting the number of times that each transition is made in the Viterbi alignments
and normalising. The final model is renamed \texttt{dhmm} and stored in
the directory \texttt{hmm0}.
When building discrete HMMs, it is important to floor the discrete
probabilites so that no symbol has a zero probability. This is
achieved using the \texttt{-w} option which specifies a floor value
as a multiple of a global constant called \texttt{MINMIX} whose
value is $10^{-5}$.
The initialised HMM created by \htool{HInit}
can then be further refined if desired by using \htool{HRest}
to perform Baum-Welch re-estimation. It would be invoked in a similar
way to the above except that there is now no need to rename the model.
For example,
\begin{verbatim}
HRest -T 1 -w 1.0 -S train.scp -M hmm1 hmm0/dhmm
\end{verbatim}
would read in the model stored in \texttt{hmm0/dhmm} and write out a new
model of the same name to the directory \texttt{hmm1}.
\mysect{Using Discrete Models with Speech}{speechvq}
As noted in section~\ref{s:vquant}, discrete HMMs can be used to model
speech by using a vector quantiser to map continuous density vectors into
discrete symbols. A vector quantiser depends on a so-called \textit{codebook}
which defines a set of partitions of the vector space. Each partition
is represented by the mean value of the speech vectors belonging
to that partition and optionally a variance representing the spread.
Each incoming speech vector is then
matched with each partition and assigned the index corresponding
to the partition which is closest using a Mahanalobis distance metric.
In \HTK\ such a codebook can be built using the tool \htool{HQuant}. This tool
takes as input a set of continuous speech vectors, clusters them and uses
the centroid and optionally the variance of each cluster to define
the partitions. \htool{HQuant} can build both linear and tree structured
codebooks. To build a linear codebook, all training vectors are initially
placed in one cluster and the mean calculated. The mean is then perturbed
to give two means and the training vectors are partitioned according to
which mean is nearest to them. The means are then recalculated and the
data is repartitioned. At each cycle, the total distortion (i.e. total
distance between the cluster members and the mean) is recorded and repartitioning
continues until there is no significant reduction in distortion. The whole
process then repeats by perturbing the mean of the cluster with the highest
distortion. This continues until the required number of clusters have been
found.
Since all training vectors are reallocated at every cycle, this is an
expensive algorithm to compute. The maximum number of iterations within
any single cluster increment can be limited using the configuration
variable \texttt{MAXCLUSTITER}\index{maxclustiter@\texttt{MAXCLUSTITER}}
and although this can speed-up the computation
significantly, the overall training process is still computationally expensive.
Once built, vector quantisation is performed by scanning all codebook
entries and finding the nearest entry. Thus, if a large codebook is used,
the run-time VQ look-up operation can also be expensive.
As an alternative to building a linear codebook, a tree-structured codebook
can be used. The algorithm for this is essentially the same as above
except that every cluster is split at each stage so that the first cluster
is split into two, they are split into four and so on. At each stage, the
means are recorded so that when using the codebook for vector quantising
a fast binary search can be used to find the appropriate leaf cluster.
Tree-structured codebooks are much faster to build since there is no
repeated reallocation of vectors and much faster in use since only $O(\log_2 N)$
distance need to be computed where $N$ is the size of the codebook.
Unfortunately, however, tree-structured codebooks will normally incur higher
VQ distortion for a given codebook size.
When delta and acceleration coefficients are used, it is usually best
to split the data into multiple streams (see section~\ref{s:streams}.
In this case, a separate codebook is built for each stream.
As an example, the following invocation of \htool{HQuant} would
generate a linear codebook in the file \texttt{linvq} using
the data stored in the files listed in \texttt{vq.scp}.
\begin{verbatim}
HQuant -C config -s 4 -n 3 64 -n 4 16 -S vq.scp linvq
\end{verbatim}
Here the configuration file \texttt{config} specifies the \texttt{TARGETKIND}
as being \texttt{MFCC\_E\_D\_A} i.e.\ static coefficients plus deltas plus
accelerations plus energy. The \texttt{-s} options requests that this
parameterisation be split into
4 separate streams. By default, each individual codebook has 256 entries, however,
the \texttt{-n} option can be used to specify alternative sizes.
If a tree-structured codebook was wanted rather than a linear codebook,
the \texttt{-t} option would be set.
Also the default is to use Euclidean distances both for building the
codebook and for subsequent coding. Setting the \texttt{-d} option
causes a diagonal covariance Mahalanobis metric to be used and
the \texttt{-f} option causes a full covariance Mahalanobis metric
to be used.
\index{hcopy@\htool{HCopy}}
\sidefig{vqtohmm}{55}{VQ Processing}{-4}{
Once the codebook is built, normal speech vector files can be
converted to discrete files using \htool{HCopy}.
This was explained
previously in section~\ref{s:vquant}. The basic mechanism is to
add the qualifier \texttt{\_V} to the
\texttt{TARGETKIND}.\index{qualifiers!aaav@\texttt{\_V}} This causes
\htool{HParm} to append a codebook index to each constructed observation
vector. If the configuration variable \texttt{SAVEASVQ} is set true, then
the output routines in \htool{HParm} will discard the original vectors
and just save the VQ indices in a \texttt{DISCRETE} file.
Alternatively, \HTK\ will regard any speech vector with \texttt{\_V} set
as being compatible with discrete HMMs. Thus, it is not necessary
to explicitly create a database of discrete training files if
a set of continuous speech vector parameter files already exists.
Fig.~\href{f:vqtohmm} illustrates this process.
}\index{saveasvq@\texttt{SAVEASVQ}}
\index{targetkind@\texttt{TARGETKIND}}
Once the training data has been configured for discrete HMMs, the
rest of the training process is similar to that previously described.
The normal sequence is to build a set of monophone models and then
clone them to make triphones. As in continuous density systems,
state tying can be used to improve the
robustness of the parameter estimates. However, in the case of discrete HMMs,
alternative methods based on interpolation are possible. These are discussed
in section~\ref{s:psmooth}.
\mysect{Tied Mixture Systems}{tiedmix}
\index{tied-mixtures}
Discrete systems have the advantage of low run-time computation. However,
vector quantisation reduces accuracy and this can lead to poor performance.
As a intermediate between discrete and continuous, a fully tied-mixture
system can be used.
Tied-mixtures are conceptually just another example of the general parameter tying
mechanism built-in to \HTK. However, to use them effectively in
speech recognition systems a number of storage and computational
optimisations must be made. Hence, they are given special treatment in \HTK.
When specific mixtures are tied as in
\begin{verbatim}
TI "mix" {*.state[2].mix[1]}
\end{verbatim}
then a Gaussian mixture component is shared across all of the owners
of the tie. In this example, all models will share the same Gaussian
for the first mixture component of state 2. However, if the mixture
component index is missing, then all of the mixture components participating in
the tie are {\it joined} rather than tied. More specifically, the commands
\begin{verbatim}
JO 128 2.0
TI "mix" {*.state[2-4].mix}
\end{verbatim}
has the following effect. All of the mixture components in states 2 to 4 of
all models are collected into a pool. If the number of components
in the pool exceeds 128, as set by the preceding join command
\texttt{JO}\index{jo@\texttt{JO} command}, then
components with the smallest weights are removed until the pool size is exactly
128. Similarly, if the size of the initial pool is less than 128, then mixture
components are split using the same algorithm as for the Mix-Up \texttt{MU}
command.\index{mixture tying} All states then share all of the
mixture components in this pool. The new mixture weights are chosen to be proportional
to the log probability of the corresponding new mixture component mean with
respect to the original distribution for that state. The log is used here
to give a wider spread of mixture weights. All mixture weights are floored
to the value of the second argument of the \texttt{JO} command times
\texttt{MINMIX}\index{minmix@\texttt{MINMIX}}.
The net effect of the above two commands is to create a set of \texttt{tied-mixture}
HMMs\footnote{Also called {\it semi-continuous} HMMs in the the literature.}
where the same set of mixture components is shared across all states of
all models. However, the type of the HMM set so created will still be
\texttt{SHARED} and the internal representation will be the same as for
any other set of parameter tyings. To obtain the optimised representation
of the tied-mixture weights
described in section~\ref{s:tmix}, the following \htool{HHEd}
\texttt{HK}\index{hk@\texttt{HK} command} command must be issued
\begin{verbatim}
HK TIEDHS
\end{verbatim}
This will convert the internal representation to the special tied-mixture
form in which all of the tied mixtures are stored in a global table and
referenced implicitly instead
of being referenced explicitly using pointers.
Tied-mixture HMMs work best if the information relating to different sources
such as delta coefficients and energy are separated into distinct data streams.
This can be done by setting up multiple data stream HMMs from the outset.
However, it is simpler to use the
\texttt{SS}\index{ss@\texttt{SS} command}
command in \htool{HHEd} to split the data streams of the currently loaded HMM set.
Thus, for example, the command
\begin{verbatim}
SS 4
\end{verbatim}
would convert the currently loaded HMMs to use four separate data streams
rather than one. When used in the construction of tied-mixture HMMs
this is analogous to the use of multiple codebooks in discrete density
HMMs.
The procedure for building a set of tied-mixture HMMs may be summarised
as follows\index{tied-mixtures!build procedure}
\begin{enumerate}
\item Choose a {\it codebook} size for each data stream and then
decide how many Gaussian components will be needed from an initial set of
monophones to approximately fill this codebook.
For example, suppose that there are 48 three state monophones. If
codebook sizes of 128 are chosen for streams 1 and 2, and
a codebook size of 64 is chosen for stream 3 then single Gaussian
monophones would provide enough mixtures in total to fill the codebooks.
\item Train the initial set of monophones.
\item Use \htool{HHEd} to first split the HMMs into the required number of
data streams, tie
each individual stream and then convert the tied-mixture HMM set to
have the kind \texttt{TIEDHS}.
A typical script to do this for four streams would be
\begin{verbatim}
SS 4
JO 256 2.0
TI st1 {*.state[2-4].stream[1].mix}
JO 128 2.0
TI st2 {*.state[2-4].stream[2].mix}
JO 128 2.0
TI st3 {*.state[2-4].stream[3].mix}
JO 64 2.0
TI st4 {*.state[2-4].stream[4].mix}
HK TIEDHS
\end{verbatim}
\item Re-estimate the models using \htool{HERest} in the normal way.
\end{enumerate}
Once the set of retrained tied-mixture models has been produced, context
dependent models can be constructed using similar methods to those
outlined previously.
When evaluating probabilities in tied-mixture systems, it is often
sufficient to sum just the most likely mixture components since for any
particular input vector, its probability with respect to many of the Gaussian
components will be very low.\index{pruning!in tied mixtures}
\HTK\ tools recognise \texttt{TIEDHS}
HMM sets as being special in the sense that additional optimisations
are possible. When full tied-mixtures are used, then an additional layer of pruning
is applied. At each time frame, the log probability of the current observation
is computed for each mixture component. Then only those components which lie within
a threshold of the most likely component are retained. This
pruning is controlled by the \texttt{-c} option in
\htool{HRest}, \htool{HERest} and \htool{HVite}.
\mysect{Parameter Smoothing}{psmooth}
When large sets of context-dependent triphones are built using
discrete models or
tied-mixture models, under-training\index{under-training} can be a
severe problem since each
state has a large number of mixture weight parameters to estimate.
The \HTK\ tool \htool{HSmooth} allows these discrete probabilities or
mixture component weights
to be smoothed with the monophone weights using a technique called
deleted interpolation\index{deleted interpolation}.
\htool{HSmooth} is used in combination with \htool{HERest}
working in parallel mode. The training data is split
into blocks and each block is used separately to re-estimate the
HMMs. However, since \htool{HERest} is in parallel mode, it outputs a dump
file of accumulators instead of updating the models. \htool{HSmooth} is then
used in place of the second pass of \htool{HERest}. It reads in the
accumulator information from each of the blocks, performs deleted
interpolation smoothing on the accumulator values and then outputs
the re-estimated HMMs in the normal way.
\htool{HSmooth}\index{hsmooth@\htool{HSmooth}} implements a
conventional deleted interpolation scheme.
However, optimisation of the smoothing weights uses a fast
binary chop\index{binary chop} scheme
rather than the more usual Baum-Welch approach.
The algorithm for finding the optimal interpolation weights for a given
state and stream is as follows where the description is given in terms
of tied-mixture weights but the same applies to discrete probabilities.
Assume that \htool{HERest}
has been set-up to output $N$ separate blocks of accumulators.
Let $w_i^{(n)}$ be the $i$'th
mixture weight based on the accumulator blocks $1$ to $N$ but excluding
block $n$, and let $\bar{w}_i^{(n)}$ be the corresponding context
independent weight. Let $x_i^{(n)}$ be the $i$'th mixture weight
count for the deleted block $n$. The derivative of the log
likelihood of the deleted block, given the probability distribution with
weights $c_i = \lambda w_i + (1 - \lambda) \bar{w}_i$ is given by
\begin{equation}
D(\lambda) = \sum_{n=1}^N \sum_{i=1}^M x_i^{(n)}
\left[ \frac{w_i^{(n)} - \bar{w}_i^{(n)}}{
\lambda w_i^{(n)} + (1 - \lambda ) \bar{w}_i^{(n)}}
\right]
\end{equation}
Since the log likelihood is a convex function of $\lambda$, this derivative
allows the optimal value of $\lambda$ to be found by a simple
binary chop algorithm, viz.
\begin{verbatim}
function FindLambdaOpt:
if (D(0) <= 0) return 0;
if (D(1) >= 0) return = 1;
l=0; r=1;
for (k=1; k<=maxStep; k++){
m = (l+r)/2;
if (D(m) == 0) return m;
if (D(m) > 0) l=m; else r=m;
}
return m;
\end{verbatim}
\htool{HSmooth} is invoked in a similar way to \htool{HERest}.
For example, suppose that the directory \texttt{hmm2} contains a set of
accumulator files output by the first pass of \htool{HERest} running in parallel
mode using as source the HMM definitions listed in \texttt{hlist} and
stored in \texttt{hmm1/HMMDefs}. Then the command
\begin{verbatim}
HSmooth -c 4 -w 2.0 -H hmm1/HMMDefs -M hmm2 hlist hmm2/*.acc
\end{verbatim}
would generate a new smoothed HMM set in \texttt{hmm2}. Here the \texttt{-w}
option is used to set the minimum mixture component weight in any state to
twice the value of \texttt{MINMIX}\index{minmix@\texttt{MINMIX}}.
The \texttt{-c} option sets the maximum
number of iterations of the binary chop procedure to be 4.
%%% Local Variables:
%%% mode: latex
%%% TeX-master: "htkbook"
%%% End: