4.2.9 Network Decoding:
Statistical Model Pool
The statistical model pool is a file consisting of gaussian mixture
representations of acoustic models. These mixtures are based on the
features that were extracted using isip_transform in.
More information about feature extraction can be found in
Section 3.
Each model is represented by a state: a symbol at the lowest level
of the search heirarchy.
The statistical model pool is created using the isip_network_builder
utility. Once the file is created, the models must be trained through
an acoustic modeling
process known as training. This process involves exposing the
statistical model pool to a series of training data, and constantly
reestimating each state based on the training data and its transcription.
The structure of the statistical model pool is in the form of a table.
The first part of this file consists of the declaration and numbering
of each state. After the declarations, each state is defined, along with
the features associated with each state. The number of states depends on
the language, the type of phoetic units (word models, context-dependent
phones), and the total number of phonetic units that exist. For a more
detailed explanation on acoustic modeling, see
Section 5.
|
|
|