TRAINING CONTINUOUS MODELS
The HMM topology file consists of a matrix with boolean entries, each entry
indiactes whether a specific transition from state=row_number to
state=column_number is permitted in the HMMs or not. For example a
3-state HMM with no skips permitted beteen states would have a topology
file with the following entries:
4
1.0 1.0 0.0 0.0
0.0 1.0 1.0 0.0
0.0 0.0 1.0 1.0
The number 4 is total the number of sates in an HMMs. The SPHINX
automatically appends a fourth non-emitting terminating state to the 3
state HMM. The first entry of 1.0 means that a transition from state 1 to
state 1 (itself) is permitted. Accordingly, the transition matrix estimated
for any phone would have a "transition-probability" in place of this
boolean entry. Where the entry is 0.0, the corresponding transition
probability will not be estimated (will be 0).
You can either write out the topology file manually, or
use the script script make_topology.pl provided with the SPHINX package to
do this. The script needs the following arguments:
states_per_hmm : this is merely an integer specifying the
number of states per hmm
skipstate : "yes" or "no" depending on whether you
want the HMMs to have skipped state transitions
or not.
Note that the topology file
is common for all HMMs and is a single file containing the topology
definition matrix. This file also defines your model architecture and is
usually placed in the model_architecture directory. This is however
optional, but recommended. If you are running scripts from the SPHINX
training package, you will find the file created in the model_architecture
directory.
Nenhum comentário:
Postar um comentário