Once the acoustic models have been seeded with initial values, a refinement
process begins. This phase of training is referred to as reestimation because
it involves applying special algorithms to reestimate the model parameters
until convergence occurs. This generates a more accurate model by building
upon the values set during initialization. The steps within the red square in
the diagram below are the phases of training considered to be reestimation.
|
The step within the diagram labeled "Build Lattice" involves
constructing a lattice, similar to the one below, which shows the
different paths for a given set of input data. Once all the possible
paths are generated, the probability that a particular piece of data
follows a certain path to a particular output can be generated for all
possible paths. A word sequence is used to generate a sequence of
phone models, then a composite HMM is formed using the phone
sequence, which leaves a state sequence that is relabeled to create the
finished lattice.
|
