speechio.tex

隐马尔科夫模型工具箱

equally on the mel-scale across the resulting pass-band such that the lower cut-off
of the first filter is at \texttt{LOFREQ} and the upper cut-off of the last
filter is at \texttt{HIFREQ}.

If mel-scale filterbank parameters are required directly, then the target kind
should be set to \texttt{MELSPEC}\index{melspec@\texttt{MELSPEC}}.
Alternatively, log filterbank parameters can be generated by setting the target
kind to \texttt{FBANK}.  


\mysect{Vocal Tract Length Normalisation}{vtln}

A simple speaker normalisation technique can be implemented by
modifying the filterbank analysis described in the previous section.
Vocal tract length normalisation (VTLN) aims to compensate for the
fact that speakers have vocal tracts of different sizes. VTLN can be
implemented by warping the frequency axis in the filterbank analysis.
In HTK simple linear frequency warping is supported. The warping
factor~$\alpha$ is controlled by the configuration variable
\texttt{WARPFREQ}\index{melspec@\texttt{WARPFREQ}}. Here values of
$\alpha < 1.0$ correspond to a compression of the frequency axis. As
the warping would lead to some filters being placed outside the
analysis frequency range, the simple linear warping function is
modified at the upper and lower boundaries. The result is that the
lower boundary frequency of the analysis
(\texttt{LOFREQ}\index{melspec@\texttt{LOFREQ}}) and the upper
boundary frequency (\texttt{HIFREQ}\index{melspec@\texttt{HIFREQ}})
are always mapped to themselves. The regions in which the warping
function deviates from the linear warping with factor~$\alpha$ are
controlled with the two configuration variables
(\texttt{WARPLCUTOFF}\index{melspec@\texttt{WARPLCUTOFF}}) and
(\texttt{WARPUCUTOFF}\index{melspec@\texttt{WARPUCUTOFF}}).
Figure~\href{f:vtlnpiecewise} shows the overall shape of the resulting
piece-wise linear warping functions.

\centrefig{vtlnpiecewise}{60}{Frequency Warping}

The warping factor~$\alpha$ can for example be found using a search
procedure that compares likelihoods at different warping factors. A
typical procedure would involve recognising an utterance with
$\alpha=1.0$ and then performing forced alignment of the hypothesis
for all warping factors in the range $0.8 - 1.2$. The factor that
gives the highest likelihood is selected as the final warping factor.
Instead of estimating a separate warping factor for each utterance,
large units can be used by for example estimating only one~$\alpha$
per speaker. 

Vocal tract length normalisation can be applied in testing as well as
in training the acoustic models.

\mysect{Cepstral Features}{cepstrum}

Most often, however, cepstral parameters are required
and these are indicated by setting the target kind to \texttt{MFCC} standing
for Mel-Frequency Cepstral Coefficients (MFCCs).  These are calculated from the
log filterbank amplitudes $\{m_j\}$ using the Discrete Cosine Transform
\hequation{
   c_i = \sqrt{\frac{2}{N}} \sum_{j=1}^N m_j \cos \left( \frac{\pi i}{N}(j-0.5) \right)
}{dct}
where $N$ is the number of filterbank channels set by the configuration
parameter \texttt{NUMCHANS}\index{numchans@\texttt{NUMCHANS}}.  The required
number of cepstral coefficients is set by
\texttt{NUMCEPS}\index{numceps@\texttt{NUMCEPS}} as in the linear prediction
case.  Liftering can also be applied to MFCCs using the
\texttt{CEPLIFTER}\index{ceplifter@\texttt{CEPLIFTER}} configuration parameter
(see equation~\ref{e:ceplifter}).

MFCCs are the parameterisation of choice for many speech recognition applications.
They give good discrimination and lend themselves to a number of manipulations.
In particular, the effect of inserting a transmission channel on the input
speech is to multiply the speech spectrum by the channel transfer function.
In the log cepstral domain, this multiplication becomes a simple addition which
can be removed by subtracting the cepstral mean from all input vectors.
In practice, of course, the mean has to be estimated over a limited amount
of speech data so the subtraction will not be perfect.  Nevertheless, this
simple technique is very effective in practice where it
compensates for long-term spectral effects such as those caused by different
microphones and audio channels.  To perform this
so-called \textit{Cepstral Mean Normalisation} (CMN) in \HTK\, it is only necessary
to add the \texttt{\_Z}\index{qualifiers!aaaz@\texttt{\_Z}} qualifier to the 
target parameter kind.  The mean is estimated by computing the average of
each cepstral parameter across each input speech file.  Since this cannot be done
with live audio, cepstral mean compensation is not supported for this case.
\index{cepstral mean normalisation}

In addition to the mean normalisation the variance of the data can be
normalised. For improved robustness both mean and variance of the data
should be calculated on a larger units (e.g.\ on all the data from a
speaker instead of just on a single utterance). To use
speaker-/cluster-based normalisation the mean and variance estimates
are computed offline before the actual recognition and stored in
separate files (two files per cluster). The configuration variables
\texttt{CMEANDIR}\index{numchans@\texttt{CMEANDIR}} and
\texttt{VARSCALEDIR}\index{numchans@\texttt{VARSCALEDIR}} point to the
directories where these files are stored. To find the actual filename
a second set of variables
(\texttt{CMEANMASK}\index{numchans@\texttt{CMEANMASK}} and
\texttt{VARSCALEMASK}\index{numchans@\texttt{VARSCALEMASK}}) has to be
specified. These masks are regular expressions in which you can use
the special characters \texttt{?}, \texttt{*} and \texttt{\%}. The
appropriate mask is matched against the filename of the file to be
recognised and the substring that was matched against the \texttt{\%}
characters is used as the filename of the normalisation file. An
example config setting is:

\begin{verbatim}
CMEANDIR     = /data/eval01/plp/cmn
CMEANMASK    = %%%%%%%%%%_*
VARSCALEDIR  = /data/eval01/plp/cvn
VARSCALEMASK = %%%%%%%%%%_*
VARSCALEFN   = /data/eval01/plp/globvar
\end{verbatim}

So, if the file \verb|sw1-4930-B_4930Bx-sw1_000126_000439.plp| is to be
recognised then the normalisation estimates would be loaded from the
following files:

\begin{verbatim}
/data/eval01/plp/cmn/sw1-4930-B
/data/eval01/plp/cvn/sw1-4930-B
\end{verbatim}

The file specified by
\texttt{VARSCALEFN}\index{numchans@\texttt{VARSCALEFN}} contains the
global target variance vector, i.e. the variance of the data is first
normalised to 1.0 based on the estimate in the appropriate file in
\texttt{VARSCALEDIR}\index{numchans@\texttt{VARSCALEDIR}} and then
scaled to the target variance given in
\texttt{VARSCALEFN}\index{numchans@\texttt{VARSCALEFN}}.

The format of the files is very simple and each of them just contains
one vector. Note that in the case of the cepstral mean only the static
coefficients will be normalised. A cmn file could for example look like:

\begin{verbatim}
<CEPSNORM> <PLP_0>
<MEAN> 13
-10.285290 -9.484871 -6.454639 ...
\end{verbatim}


The cepstral variance normalised always applies to the full
observation vector after all qualifiers like delta and acceleration
coefficients have been added, e.g.:

\begin{verbatim}
<CEPSNORM> <PLP_D_A_Z_0>
<VARIANCE> 39
33.543018 31.241779 36.076199 ...
\end{verbatim}

The global variance vector will always have the same number of
dimensions as the cvn vector, e.g.:

\begin{verbatim}
<VARSCALE> 39
 2.974308e+01 4.143743e+01 3.819999e+01 ...
\end{verbatim}

These estimates can be generated using \htool{HCompV}. See the
reference section for details.

 
\mysect{Perceptual Linear Prediction}{plp}

An alternative to the Mel-Frequency Cepstral Coefficients is the use
of Perceptual Linear Prediction (PLP) coefficients.

As implemented in HTK the PLP feature extraction is based on the
standard mel-frequency filterbank (possibly warped). The mel
filterbank coefficients are weighted by an equal-loudness curve and
then compressed by taking the cubic root.\footnote{the degree of
  compression can be controlled by setting the configuration parameter
  \texttt{COMPRESSFACT}\index{enormalise@\texttt{COMPRESSFACT}} which
  is the power to which the amplitudes are raised and defaults to
  0.33)} From the resulting auditory spectrum LP coefficents are
estimated which are then converted to cepstral coefficents in the
normal way (see above).


\mysect{Energy Measures}{energy}

\index{speech input!energy measures} 
To augment the spectral parameters derived from linear prediction or
mel-filterbank analysis, an energy term can be appended by including the
qualifier \texttt{\_E}\index{qualifiers!aaae@\texttt{\_E}} in the target kind.
The energy is computed as the log of the signal energy, that is, for speech
samples $\{s_n, n=1,N \}$
\hequation{
   E = log \sum_{n=1}^N s_n^2
}{logenergy}

This log energy measure can be normalised to the range $-E_{min}..1.0$ by
setting the Boolean configuration parameter
\texttt{ENORMALISE}\index{enormalise@\texttt{ENORMALISE}} to true (default
setting).  This
normalisation is implemented by subtracting the maximum value of $E$ in the
utterance and adding $1.0$. 
Note that energy normalisation is incompatible with live audio 
input and in such circumstances the configuration variable \texttt{ENORMALISE} 
should be explicitly set false.
The lowest energy in the utterance can be clamped using the configuration
parameter
\texttt{SILFLOOR}\index{silfloor@\texttt{SILFLOOR}} which gives the ratio
between the maximum and minimum energies in the utterance in dB. Its default
value is 50dB. 
Finally, the overall log energy can be arbitrarily scaled by the value of the 
configuration parameter \texttt{ESCALE}\index{escale@\texttt{ESCALE}} whose 
default is $0.1$. \index{silence floor}

When calculating energy for LPC-derived parameterisations, the default is to
use the zero-th delay autocorrelation coefficient ($r_0$).  However, this means
that the energy is calculated after windowing and pre-emphasis.  If the
configuration parameter \texttt{RAWENERGY}\index{rawenergy@\texttt{RAWENERGY}}
is set true, however, then energy is calculated separately before any windowing
or pre-emphasis regardless of the requested parameterisation\footnote{ In any
event, setting the compatibility variable \texttt{V1COMPAT} to true in
\htool{HPARM} will ensure that the calculation of energy is compatible with
that computed by the Version 1 tool \htool{HCode}.  }.

In addition to, or in place of, the log energy, the qualifier
\texttt{\_O}\index{qualifiers!aaao@\texttt{\_O}} can be added to a target kind
to indicate that the 0'th cepstral parameter $C_0$ is to be appended.  This
qualifier is only valid if the target kind is \texttt{MFCC}.  Unlike earlier
versions of \HTK\, scaling factors set by the configuration variable
\texttt{ESCALE} are not applied to $C_0$\footnote{ Unless \texttt{V1COMPAT} is
set to true.  }.

\mysect{Delta, Acceleration and Third Differential Coefficients}{delta}

\index{speech input!dynamic coefficents}
The performance of a speech recognition system can be greatly enhanced by
adding time derivatives to the basic static parameters.  In \HTK, these are
indicated by attaching qualifiers to the basic parameter kind.  The qualifier
\texttt{\_D} indicates that first order regression coefficients (referred to as
delta coefficients) are appended, the qualifier
\texttt{\_A}\index{qualifiers!aaaa@\texttt{\_A}} indicates that second order
regression coefficients (referred to as acceleration coefficients) and 
 the qualifier
\texttt{\_T}\index{qualifiers!aaaa@\texttt{\_T}} indicates that third order
regression coefficients (referred to as third differential coefficients) are
appended. The \texttt{\_A} qualifier cannot be used without also using the
\texttt{\_D}\index{qualifiers!aaad@\texttt{\_D}} qualifier. Similarly
the \texttt{\_T} qualifier cannot be used without also using the
\texttt{\_D} and \texttt{\_A} qualifiers.

The delta coefficients\index{delta coefficients} are computed using the
following regression formula\index{regression formula}
\hequation{
   d_t = \frac{ \sum_{\theta =1}^\Theta \theta(c_{t+\theta} - c_{t-\theta}) }{
                2 \sum_{\theta = 1}^\Theta \theta^2 }
}{deltas}
where $d_t$ is a delta coefficient at time $t$ computed in terms of the
corresponding static coefficients $c_{t-\Theta}$ to $c_{t+\Theta}$.  The value
of $\Theta$ is set using the configuration parameter
\texttt{DELTAWINDOW}\index{deltawindow@\texttt{DELTAWINDOW}}.  The same formula
is applied to the delta coefficients to obtain acceleration coefficients except
that in this case the window size is set by
\texttt{ACCWINDOW}\index{accwindow@\texttt{ACCWINDOW}}. Similarly
the third differentials use \texttt{THIRDWINDOW}. Since
equation~\ref{e:deltas} relies on past and future speech parameter values, 
some modification is needed at the beginning and end of the speech.  The
default behaviour is to replicate the first or last vector as needed to fill