hquant.tex

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

%/* ----------------------------------------------------------- */
%/*                                                             */
%/*                          ___                                */
%/*                       |_| | |_/   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 11/11/95
%

\newpage
\mysect{HQuant}{HQuant}

\mysubsect{Function}{HQuant-Function}

\index{hquant@\htool{HQuant}|(}
This program will construct a \HTK\ format VQ table consisting of one or
more codebooks, each corresponding to an independent data stream.  
A codebook is a collection of indexed reference vectors that are intended to
represent the structure of the training data. Ideally, every compact
cluster of points in the training data is represented by one reference
vector.  VQ
tables are used directly when building systems based on Discrete
Probability HMMs.  In this case, the continuous-valued speech vectors
in each stream are replaced by the index of the closest reference
vector in each corresponding codebook.
Codebooks can also be used to attach VQ indices to
continuous observations.  A typical use of this is to preselect
Gaussians in probability computations.  More information on the
use of VQ tables is given in section~\ref{s:vquant}.

Codebook construction consists of finding clusters in the training data,
assigning a unique reference vector (the cluster centroid) to each,
and storing the resultant reference vectors and associated codebook data
in a VQ table. \htool{HQuant} uses a top-down clustering process whereby
clusters are iteratively split until the desired number of clusters are
found. The optimal codebook size (number of reference vectors) depends
on the structure and amount of the training data, but a value of 256 is
commonly used.

\htool{HQuant} can construct both linear (i.e.\ flat) and tree-structured
(i.e.\ binary) codebooks.  Linear codebooks can have lower quantisation
errors but tree-structured codebooks have $\log_2 N$ access times
compared to $N$ for the linear case.  The distance metric can either be
Euclidean, diagonal covariance Mahalanobis or full covariance
Mahalanobis.

\mysubsect{VQ Codebook Format}{HQuant-VQ Codebook Format}

Externally, a VQ table is stored in a text file consisting of a header
followed by a sequence of entries representing each tree node.  One tree
is built for each stream and linear codebooks are represented by a tree
in which there are only right branches.

The header consists of a magic number followed by the covariance kind,
the number of following nodes, the number of streams and the width
of each stream.
{\sf
\begin{tabbing}
++++ \= ++++++++ \=  \kill
\>  header =\> magic type covkind numNodes numS swidth1 swidth2  \ldots
\end{tabbing}
}
\noindent
where {\sf magic} is a magic number which is usually the
code for the parameter kind of the data.  The {\sf type}
defines the type of codebook
{\sf
\begin{tabbing}
++++ \= ++++++++ \=  \kill
\>  type =\> linear (0) , binary tree-structured (1)
\end{tabbing}
}
\noindent
The covariance kind determines the type of distance metric to be
used
{\sf
\begin{tabbing}
++++ \= ++++++++ \=  \kill
\>  covkind =\> diagonal covariance (1), full covariance (2), euclidean (5)
\end{tabbing}
}
\noindent
Within the file, these covariances are stored in inverse form.

Each node entry has the following form
{\sf
\begin{tabbing}
++++ \= ++++++++ \=  \kill
\>  node-entry =\> stream vqidx nodeId leftId rightId \\
\> \> mean-vector \\
\> \> [inverse-covariance-matrix $|$ inverse-variance-vector]
\end{tabbing}
}
\noindent
{\sf Stream} is the stream index for this entry. {\sf Vqidx}
is the VQ index corresponding to this entry.  This is the number that
appears in vector quantised speech files.  In tree-structured code-books,
it is zero for non-terminal nodes. Every node has a unique
integer identifier (distinct from the VQ index) given by {\sf nodeId}.
The left and right daughter of the current node are given by
{\sf leftId} and {\sf rightId}.  In a linear codebook, the left identifier
is always zero.

Some examples of VQ tables are given in Chapter~\ref{c:discmods}.

\mysubsect{Use}{HQuant-Use}

\htool{HQuant} is invoked via the command line

\begin{verbatim}
   HQuant [options] vqFile trainFiles ...
\end{verbatim}
where \texttt{vqFile} is the name of the output VQ table file.
The effect of this command
is to read in training data from each \texttt{trainFile}, cluster the
data and write the final cluster centres into the VQ table file.

The list of training files can be stored in a script file if required.
Furthermore, the data used for training the codebook
can be limited to that corresponding to a specified label.  This can be
used, for example, to train phone specific codebooks.  When constructing
a linear codebook, the maximum number of iterations per cluster can be
limited by setting the configuration variable \texttt{MAXCLUSTITER}.
The minimum number of samples in any one cluster can be set using the
 configuration variable \texttt{MINCLUSTSIZE}.

The detailed operation of \htool{HQuant} is controlled by the following
command line options
\begin{optlist}
   \ttitem{-d} Use a diagonal-covariance Mahalonobis distance metric for
 clustering (default is to use a Euclidean distance metric).

   \ttitem{-f} Use a full-covariance Mahalonobis distance metric for
 clustering (default is to use a Euclidean distance metric).

   \ttitem{-g} Output the global covariance to a codebook.  Normally,
     covariances are computed individually for each cluster using
     the data in that cluster.  This option computes a global covariance
     across all the clusters.

  \ttitem{-l s} The string {\tt s} must be the name of a
      segment label.  When this option is used, \htool{HQuant} searches
      through all of the training files and uses only the speech
      frames from segments with the given label.  When this option is not 
      used, \htool{HQuant} uses all of the data in each training file.

  \ttitem{-n S N} Set size of codebook for stream \texttt{S} 
       to \texttt{N} (default 256).
   If tree-structured codebooks are required then \texttt{N} 
   must be a power of 2.
  
  \ttitem{-s N} Set number of streams to \texttt{N} (default 1).
    Unless the \texttt{-w} option is used, the width of each stream
    is set automatically depending on the size and parameter kind of the 
    training data.

  \ttitem{-t}  Create tree-structured codebooks (default linear).
  
  \ttitem{-w S N} Set width of stream \texttt{S} to \texttt{N}.
  This option overrides the default decomposition that \HTK\ normally
  uses to divide a parameter file into streams.  If this option is used,
  it must be repeated for each individual stream.
  
\stdoptF
\stdoptG
\stdoptI
\stdoptL
\stdoptX
\end{optlist}
\stdopts{HQuant}
\mysubsect{Tracing}{HQuant-Tracing}

\htool{HQuant} supports the following trace options where each
trace flag is given using an octal base
\begin{optlist}
   \ttitem{00001} basic progress reporting.
   \ttitem{00002} dump global mean and covariance
   \ttitem{00004} trace data loading.
   \ttitem{00010} list label segments.
   \ttitem{00020} dump clusters.
   \ttitem{00040} dump VQ table.
\end{optlist}
Trace flags are set using the \texttt{-T} option or the  \texttt{TRACE} 
configuration variable.

\index{hquant@\htool{HQuant}|)}


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