.#htkoview.tex.1.3

该压缩包为最新版htk的源代码,htk是现在比较流行的语音处理软件,请有兴趣的朋友下载使用

The philosophy of system construction in \HTK\  is that HMMs should be
\index{HMM!build philosophy}
refined incrementally.  Thus, a typical progression is to start with a
simple set of single Gaussian context-independent phone models and then
iteratively refine them by expanding them to include context-dependency 
and use multiple mixture component Gaussian distributions.
The tool \htool{HHEd}\index{hhed@\htool{HHEd}} is a 
HMM definition editor which will clone models\index{HMM!editor}
into context-dependent sets, apply a variety of parameter tyings
and increment the number of mixture components in specified distributions.
The usual process is to modify a set of HMMs in stages using \htool{HHEd}
and then 
re-estimate the parameters of the modified set using \htool{HERest}
after each stage.  
To improve performance for specific speakers the tools 
\htool{HERest}\index{herest@\htool{HERest}} 
and \htool{HVite}\index{hvite@\htool{HVite}} can be 
used to adapt HMMs to better model the characteristics of particular 
speakers using a small amount of training or adaptation data.
The end result of which is a speaker adapted system.

The single biggest problem in building context-dependent 
HMM systems is always data
insufficiency.  The more complex the model set, the more data is 
needed to make robust estimates of its parameters, and since data is
usually limited, a balance must be struck between complexity and 
the available data.
For continuous density systems, this balance is achieved by 
tying parameters together as mentioned above.  Parameter tying
allows data to be pooled so that the shared parameters can be robustly
estimated. 
In addition to continuous density systems, \HTK\ also supports
fully tied mixture systems and discrete probability systems.  In these
cases, the data insufficiency problem is usually addressed by smoothing
the distributions and the 
tool \htool{HSmooth}\index{hsmooth@\htool{HSmooth}} is used for this.

\subsection{Recognition Tools}

\HTK\ provides a single recognition 
tool\index{recognition!tools} called \htool{HVite}\index{hvite@\htool{HVite}}
which uses the token passing algorithm described in the previous
chapter to perform Viterbi-based speech recognition.  \htool{HVite}
takes as input a network describing the allowable word sequences,
a dictionary defining how each word is pronounced and a set of HMMs.
It operates by converting the word network to a phone network and
then attaching the appropriate HMM definition to each phone instance.
Recognition can then be performed on either a list of stored speech
files or on direct audio input.  As noted at the end of the last
chapter, \htool{HVite} can support cross-word triphones and it can
run with multiple tokens to generate
lattices containing multiple hypotheses.  It can also be configured
to rescore lattices and perform forced alignments.

The word networks needed to drive \htool{HVite} are usually either
simple word loops in which any word can follow any other word or they
are directed graphs representing a finite-state task grammar.  In the
former case, bigram probabilities are normally attached to the word
transitions.  Word networks are stored using
the \HTK\ standard lattice format\index{lattice!format}.  
This is a text-based format and\index{standard lattice format}\index{SLF}
hence word networks can be created directly using a text-editor.
However, this is rather tedious and hence \HTK\ provides two
tools to assist in creating word networks.  Firstly, \htool{HBuild} 
allows sub-networks to be created and used within higher level networks.
Hence, although the same low level notation is used, much duplication
is avoided.  Also, \htool{HBuild}\index{hbuild@\htool{HBuild}} can be used to generate word loops
and it can also read in a backed-off bigram language model and 
modify the word loop transitions to incorporate the bigram probabilities.
Note that the label statistics tool \htool{HLStats} mentioned earlier
can be used to generate a backed-off bigram language model.

As an alternative to specifying a word network directly, a higher
level grammar notation can be used.  This notation is based on
the Extended Backus Naur Form (EBNF\index{EBNF}) used in compiler specification and
it is compatible with the grammar specification language used in 
earlier versions of \HTK.  The 
tool \htool{HParse}\index{hparse@\htool{HParse}} is supplied
to convert this notation into the equivalent word network.

Whichever method is chosen to generate a word network, it is useful
to be able to see examples of the \textit{language} that it defines.
The tool \htool{HSGen}\index{hsgen@\htool{HSGen}} is 
provided to do this.  It takes as input
a network and then randomly traverses the network outputting word
strings.  These strings can then be inspected to ensure that they
correspond to what is required.  \htool{HSGen} can also compute
the empirical perplexity of the task.

Finally, the construction of large dictionaries can involve merging
several sources and performing a variety of transformations on each
sources.  The dictionary management 
tool \htool{HDMan}\index{hdman@\htool{HDMan}} is supplied
to assist with this process.

\subsection{Analysis Tool}

\index{results analysis}
Once the HMM-based recogniser has been built, it is necessary
to evaluate its performance.  This is usually done by using it
to transcribe some pre-recorded test sentences and match the
recogniser output with the correct reference transcriptions.
This comparison is performed by a tool called 
\htool{HResults} which uses dynamic programming to align the two transcriptions
and then count substitution, deletion and insertion errors.
Options are provided to ensure that the
algorithms and output formats used 
by \htool{HResults}\index{hresults@\htool{HResults}} are compatible
with those used by the US National Institute of Standards and Technology
(NIST).
As well as global performance measures,
\htool{HResults} can also provide speaker-by-speaker breakdowns,
confusion matrices and time-aligned transcriptions.  For word spotting
applications, it can also compute \textit{Figure of Merit} (FOM) scores
and \textit{Receiver Operating Curve} (ROC) information.
\index{NIST}\index{FOM}\index{Figure of Merit}


\mysect{Whats New In Version 3.3}{whatsnew}

This \index{new features!in Version 3.3} section lists the new
features in \HTK\ Version 3.3 compared to the preceding Version~3.3.

\begin{enumerate}
\item \htool{HERest} now incorporates the adaptation transform generation
  that was previously performed in \htool{HEAdapt}. The range of linear
  transformations and the ability to combine transforms hierarchically
  has now been included. The system also now supports adaptive training
  with contrained MLLR transforms.

\item Many other smaller changes and bug fixes have been integrated.
\end{enumerate}

\mysubsect{New In Version 3.2}{}

This \index{new features!in Version 3.2} section lists the new
features in \HTK\ Version 3.2 compared to the preceding Version~3.1.

\begin{enumerate}
\item The \htool{HLM} toolkit has been incorporated into HTK. It
  supports the training and testing of word or class-based n-gram
  language models. 
\item \htool{HPARM} supports global feature space transforms.
\item \htool{HPARM} now supports third differentials
  ($\Delta\Delta\Delta$ parameters).
\item A new tool named \htool{HLRescore} offers support for a number
  of lattice post-processing operations such as lattice pruning,
  finding the 1-best path in a lattice and language model expansion of
  lattices.
\item \htool{HERest} supports 2-model re-estimation which allows the
  use of a separate alignment model set in the Baum-Welch
  re-estimation.
\item The initialisation of the decision-tree state clustering in
  \htool{HHEd} has been improved.
\item \htool{HHEd} supports a number of new commands related to
  variance flooring and decreasing the number of mixtures.
\item A major bug in the estimation of block-diagonal MLLR transforms
  has been fixed.
\item Many other smaller changes and bug fixes have been integrated.
\end{enumerate}

\subsection{New In Version 3.1}{}

This \index{new features!in Version 3.1} section lists the new
features in \HTK\ Version 3.1 compared to the preceding Version~3.0
which was functionally equivalent to Version~2.2.

\begin{enumerate}

\item \htool{HPARM} supports Perceptual Linear Prediction (PLP) feature
  extraction.

\item \htool{HPARM} supports Vocal Tract Length Normalisation (VTLN)
  by warping the frequency axis in the filterbank analysis.

\item \htool{HPARM} supports variance scaling.

\item \htool{HPARM} supports cluster-based cepstral mean and variance
  normalisation. 

\item All tools support an extended filename syntax that can be used
  to deal with unsegmented data more easily.

\end{enumerate}


\subsection{New In Version 2.2}{}

This section lists the new features and refinements in \HTK\ Version
2.2 compared to the preceding Version 2.1.
 
\begin{enumerate}

\item Speaker adaptation is now supported via the \htool{HEAdapt} and 
\htool{HVite} tools, which adapt a current set of models to a new speaker and/or
environment.
\begin{itemize}

\item \htool{HEAdapt} performs offline supervised adaptation using maximum 
likelihood linear regression (MLLR) and/or maximum a-posteriori (MAP) adaptation.

\item \htool{HVite} performs unsupervised adaptation using just MLLR.

\end{itemize}

Both tools can be used in a static mode, where all the
data is presented prior to any adaptation, or in an incremental fashion.

\item Improved support for PC WAV files\\
In addition to 16-bit PCM linear, HTK can now read  
\begin{itemize}

\item 8-bit CCITT mu-law

\item 8-bit CCITT a-law

\item 8-bit PCM linear

\end{itemize}

\end{enumerate}

\subsection{Features Added To Version 2.1}{}

For the benefit of users of earlier versions of \HTK\, this 
\index{new features!in Version 2.1} section lists the main changes 
in \HTK\ Version 2.1 compared to the preceding Version 2.0.

\begin{enumerate}

\item The speech input handling has been partially re-designed and a new 
energy-based speech/silence detector has been incorporated into \htool{HParm}.
The detector is robust yet flexible and can be configured through a number of
configuration variables. Speech/silence detection can now be performed on
waveform files. The calibration of speech/silence detector parameters is now
accomplished by asking the user to speak an arbitrary sentence.

\item \htool{HParm} now allows random noise signal to be added to waveform
data via the configuration parameter \texttt{ADDDITHER}. This prevents
numerical overflows which can occur with artificially created waveform data
under some coding schemes.

\item \htool{HNet} has been optimised for more efficient operation when
performing forced alignments of utterances using \htool{HVite}. Further
network optimisations tailored to biphone/triphone-based phone recognition 
have also been incorporated.

\item \htool{HVite} can now produce partial recognition hypothesis even when 
no tokens survive to the end of the network. This is accomplished by setting
the \htool{HRec} configuration parameter \texttt{FORCEOUT} to true.

\item Dictionary support has been extended to allow pronunciation probabilities
to be associated with different pronunciations of the same word. At the same
time, \htool{HVite} now allows the use of a pronunciation scale factor during
recognition.

\item \HTK\ now provides consistent support for reading and writing of \HTK\ 
binary files (waveforms, binary MMFs, binary SLFs, \htool{HERest} accumulators)
across different machine architectures incorporating automatic byte swapping.
By default, all binary data files handled by the tools are now written/read in
big-endian (\texttt{NONVAX}) byte order. The default behavior can be changed
via the configuration parameters \texttt{NATURALREADORDER} and
\texttt{NATURALWRITEORDER}.

\item \htool{HWave} supports the reading of waveforms in Microsoft WAVE file
format.

\item \htool{HAudio} allows key-press control of live audio input.

\end{enumerate}



%%% Local Variables: 
%%% mode: latex
%%% TeX-master: "htkbook"
%%% End: