speechio.tex

隐马尔科夫模型工具箱

the regression window.

In older version 1.5 of \HTK\ and earlier, this end-effect problem was solved
by using simple 
first order differences at the start and end of the speech, that is
\begin{equation}
   d_t = c_{t+1} - c_t,\;\;\; t<\Theta
\end{equation}
and
\begin{equation}
   d_t = c_t - c_{t-1}, \;\;\; t \geq T-\Theta
\end{equation}
where $T$ is the length of the data file.  If required, this older behaviour
can be restored by setting the configuration variable 
\texttt{V1COMPAT}\index{v1compat@\texttt{V1COMPAT}}
to true in \htool{HParm}.

For some purposes, it is useful to use simple differences throughout.  This
can be achieved by setting the configuration 
variable \texttt{SIMPLEDIFFS}\index{simplediffs@\texttt{SIMPLEDIFFS}}
to true in \htool{HParm}.  In this case, just the end-points of the delta window
are used, i.e.
\hequation{
   d_t = \frac{ (c_{t+\Theta} - c_{t-\Theta}) }{
                2 \Theta}
}{simdiffs}
\index{simple differences}

When delta and acceleration coefficients are requested, they are computed for
all static parameters including energy if present.  In some applications, the
absolute energy is not useful but time derivatives of the energy may be.  By
including the \texttt{\_E} qualifier together with the
\texttt{\_N}\index{qualifiers!aaan@\texttt{\_N}} qualifier, the absolute energy
is suppressed leaving just the delta and acceleration coefficients of the
energy.

\mysect{Storage of Parameter Files}{parmstore}

Whereas \HTK\ can handle waveform data in a variety of file formats,
all parameterised speech data is stored externally in either native
\HTK\ format data files or Entropic Esignal format files.
Entropic ESPS format is no longer supported directly, but input and output
filters can be used to convert ESPS to Esignal format on input and
Esignal to ESPS on output.

\subsection{\HTK\ Format Parameter Files}

\HTK\ format files consist of a contiguous sequence of \textit{samples}
preceded by a header.  Each sample is a vector of either 2-byte integers or
4-byte floats.  2-byte integers are used for compressed forms as described
below and for vector quantised data as described later in
section~\ref{s:vquant}.  \HTK\ format data files can also be used to store
speech waveforms as described in section~\ref{s:waveform}.  \index{file
formats!HTK}

The \HTK\ file format header is 12 bytes long and contains the following data 
\begin{tabbing}
++ \= +++++++++ \= \kill
\>\texttt{nSamples}\>-- number of samples in file (4-byte integer) \\
\>\texttt{sampPeriod}\>-- sample period in 100ns units (4-byte integer) \\
\>\texttt{sampSize}\>-- number of bytes per sample (2-byte integer) \\
\>\texttt{parmKind}\>-- a code indicating the sample kind (2-byte integer)
\end{tabbing}
The parameter kind\index{parameter kind} consists of a 6 bit
code representing the basic parameter kind plus additional bits for
each of the possible qualifiers\index{qualifiers}.  The basic parameter kind codes are
\begin{tabbing}
++++\= +++ \= ++++++++ \=   \kill
\>0 \> \texttt{WAVEFORM} \> sampled waveform \\
\>1 \> \texttt{LPC} \> linear prediction filter coefficients \\
\>2 \> \texttt{LPREFC} \> linear prediction reflection coefficients \\
\>3 \> \texttt{LPCEPSTRA} \> LPC cepstral coefficients \\
\>4 \> \texttt{LPDELCEP}  \> LPC cepstra plus delta coefficients \\
\>5 \> \texttt{IREFC}       \> LPC reflection coef in 16 bit integer format  \\
\>6 \> \texttt{MFCC}      \> mel-frequency cepstral coefficients \\
\>7 \> \texttt{FBANK}     \> log mel-filter bank channel outputs \\
\>8 \> \texttt{MELSPEC}     \> linear mel-filter bank channel outputs \\
\>9 \> \texttt{USER}      \> user defined sample kind \\
\>10 \> \texttt{DISCRETE} \> vector quantised data \\
\end{tabbing}
and the bit-encoding for the qualifiers (in octal) is 
\begin{tabbing}
++++\= +++ \= ++++++++ \=   \kill
\>\texttt{\_E} \> 000100 \> has energy \\
\>\texttt{\_N} \> 000200 \> absolute energy suppressed \\
\>\texttt{\_D} \> 000400 \> has delta coefficients \\
\>\texttt{\_A} \> 001000 \> has acceleration coefficients\\
\>\texttt{\_C} \> 002000 \> is compressed \\
\>\texttt{\_Z} \> 004000 \> has zero mean static coef. \\
\>\texttt{\_K} \> 010000 \> has CRC checksum \\
\>\texttt{\_O} \> 020000 \> has 0'th cepstral coef. \\
\end{tabbing}\index{qualifiers!codes}
The \texttt{\_A} qualifier can only be specified when \texttt{\_D}
is also specified.
The \texttt{\_N} qualifier is only valid when both energy and delta
coefficients are present.  
The sample kind \texttt{LPDELCEP} is identical to \texttt{LPCEPSTRA\_D}
and is retained for compatibility with older versions of \HTK.
The \texttt{\_C}\index{qualifiers!aaac@\texttt{\_C}} and 
\texttt{\_K}\index{qualifiers!aaak@\texttt{\_K}} only exist in external files.  Compressed
files are always decompressed on loading and any attached CRC 
is checked and removed.  An external file can contain both an energy
term and a 0'th order cepstral coefficient.  These may be retained
on loading but normally one or the other is discarded\footnote{
Some applications may require the 0'th order cepstral coefficient
in order to recover the filterbank coefficients from the cepstral
coefficients.}.

\putfig{HTKFormat}{130}{Parameter Vector Layout in \HTK\ Format Files}

All parameterised forms of \HTK\ data files consist of a sequence of vectors.
Each vector is organised as shown by the examples in Fig~\href{f:HTKFormat}
where various different qualified forms are listed.  As can be seen, an energy
value if present immediately follows the base coefficients.  If delta
coefficients are added, these follow the base coefficients and energy value.
Note that the base form \texttt{LPC} is used in this figure only as an example,
the same layout applies to all base sample kinds.  If the 0'th order cepstral
coefficient is included as well as energy then it is inserted immediately
before the energy coefficient, otherwise it replaces it.

For external storage of speech parameter files, two compression methods are
provided.  For LP coding only, the \texttt{IREFC} parameter kind exploits the
fact that the reflection coefficients are bounded by $\pm 1$ and hence they can
be stored as scaled integers such that $+1.0$ is stored as $32767$ and $-1.0$
is stored as $-32767$.  For other types of parameterisation, a more general
compression facility indicated by the
\texttt{\_C}\index{qualifiers!aaac@\texttt{\_C}} qualifier is used.  
\HTK\ compressed parameter files consist of a set of compressed parameter 
vectors stored as shorts such that for parameter $x$
\begin{eqnarray}
x_{short} & = & A*x_{float}-B  \nonumber   
\end{eqnarray}
The coefficients $A$ and $B$ are defined as
\begin{eqnarray}
A & = & 2*I/(x_{max}-x_{min}) \nonumber\\
B & = & (x_{max}+x_{min})*I/(x_{max}-x_{min}) \nonumber
\end{eqnarray}
where $x_{max}$ is the maximum value of parameter $x$ in the whole file and
$x_{min}$ is the corresponding minimum. $I$ is the maximum range of a 2-byte
integer i.e.\ 32767.  The values of $A$ and $B$ are stored as two floating
point vectors prepended to the start of the file immediately after the header.

When a \HTK\ tool writes out a speech file to external storage, no further
signal conversions are performed.  Thus, for most purposes, the target
parameter kind specifies both the required internal representation and the form
of the written output, if any.  However, there is a distinction in the way that
the external data is actually stored.  Firstly, it can be compressed as
described above by setting the configuration parameter \texttt{SAVECOMPRESSED}
to true.  If the target kind is \texttt{LPREFC} then this compression is
implemented by converting to \texttt{IREFC} otherwise the general compression
algorithm described above is used.  Secondly, in order to avoid data corruption
problems, externally stored \HTK\ parameter files can have a cyclic redundancy
checksum appended.  This is indicated by the qualifier
\texttt{\_K}\index{qualifiers!aaak@\texttt{\_K}} and it is generated by setting
the configuration parameter \texttt{SAVEWITHCRC} to true.  The principle tool
which uses these output conversions is \htool{HCopy} (see
section~\ref{s:UseHCopy}).

\subsection{Esignal Format Parameter Files}

\index{file formats!Esignal}
The default for parameter files is native \HTK\ format.  However, \HTK\ tools
also support the Entropic Esignal format for both input and output. Esignal
replaces the Entropic ESPS file format. To ensure compatibility Entropic
provides conversion programs from ESPS to ESIG and vice versa.

To indicate that a source file is in Esignal format the configuration
variable  \texttt{SOURCEFORMAT}\index{sourceformat@\texttt{SOURCEFORMAT}}
should be set to \texttt{ESIG}.  Alternatively, 
\texttt{-F ESIG}\index{standard options!aaaf@\texttt{-F}} can be specified
as a command-line option.  
To generate Esignal format output files, the configuration variable
\texttt{TARGETFORMAT} should be set to \texttt{ESIG} or the command line option
\texttt{-O ESIG} should be set.

ESIG files consist of three parts: a preamble, a sequence of field 
specifications called the field list and a sequence of records. The preamble
and the field list together constitute the header. The preamble is purely 
ASCII. Currently it consists of 6 information items that are all terminated
by a new line.  The information in the preamble is the following:
\begin{tabbing}
++ \= +++++++++ \= \kill
\>\texttt{line 1}\>-- identification of the file format \\
\>\texttt{line 2}\>-- version of the file format\\
\>\texttt{line 3}\>-- architecture (ASCII, EDR1, EDR2, machine name)\\
\>\texttt{line 4}\>-- preamble size (48 bytes)\\
\>\texttt{line 5}\>-- total header size\\
\>\texttt{line 6}\>-- record size\\
\end{tabbing}
All ESIG files that are output by \HTK\ programs contain the following 
global fields:
\begin{description}
  \item[commandLine] the command-line used to generate the file;
  \item[recordFreq] a double value that indicates the sample frequency
        in Herz;
  \item[startTime] a double value that indicates a time at which the first 
        sample is presumed to be starting;
  \item[parmKind] a character string that indicates the full 
        type of parameters in the file, e.g: \texttt{MFCC\_E\_D}.
  \item[source\_1] if the input file was an ESIG file this field includes the
        header items in the input file.
\end{description}
After that there are field specifiers for the records. The first specifier 
is for the  basekind of the parameters, e.g: \texttt{MFCC}. Then for each 
available qualifier there are additional specifiers. Possible specifiers are:
\begin{tabbing}
++++\=   \kill
\>\texttt{zeroc}  \\
\>\texttt{energy}\\
\>\texttt{delta}\\
\>\texttt{delta\_zeroc} \\
\>\texttt{delta\_energy}\\
\>\texttt{accs}\\
\>\texttt{accs\_zeroc} \\
\>\texttt{accs\_energy}\\
\end{tabbing}\index{qualifiers!ESIG field specifiers}
The data segments of the ESIG files have exactly the same format as the
the corresponding \HTK\ files. This format was described in the previous 
section.

\HTK\ can only input parameter files that have a valid parameter kind as value
of the header field \texttt{parmKind}. If this field does not exist or if the
value of this field does not contain a valid parameter kind, the file is 
rejected. After the header has been read the file is treated as an \HTK\ file.


\mysect{Waveform File Formats}{waveform}

For reading waveform data files, \HTK\ can support a variety of different
formats and these are all briefly described in this section.  The default
speech file format is \HTK. If a different format is to be used, it can be
specified by setting the configuration parameter
\texttt{SOURCEFORMAT}\index{sourceformat@\texttt{SOURCEFORMAT}}.  However,
since file formats need to be changed often, they can also be set individually
via the \texttt{-F}\index{standard options!aaaf@\texttt{-F}} command-line
option.  This over-rides any setting of the \texttt{SOURCEFORMAT} configuration
parameter.

Similarly for the output of waveforms, the format can be set using either the
configuration parameter \texttt{TARGETFORMAT} or the \texttt{-O} command-line
option.  However, for output only native \HTK\ format (\texttt{HTK}), Esignal
format (\texttt{ESIG}) and headerless (\texttt{NOHEAD}) waveform files are
supported.

The following sub-sections give a brief description of each of the waveform
file formats supported by \HTK.

\subsection{HTK File Format}

\index{file formats!HTK}
The \HTK\ file format for waveforms is identical to that described in
section~\ref{s:parmstore} above.  It consists of a 12 byte header followed
by a sequence of 2 byte integer speech samples.  For waveforms, the
\texttt{sampSize} field will be 2 and the \texttt{parmKind} field will be 0.
The \texttt{sampPeriod} field gives the sample period in 100ns units, hence for
example, it will have the value 1000 for speech files sampled at 10kHz and 625
for speech files sampled at 16kHz.

\subsection{Esignal File Format}

\index{file formats!Esignal}
The Esignal file format for waveforms is similar to that described in
section~\ref{s:parmstore} above with the following exceptions. When reading an