Digital Signal Processing Reference
In-Depth Information
The following, dynamic network aggregates the time-varying local
distances
d
to a temporally varying distance-vector
g.
In case of
continuous recognition, hypotheses are selected from the search space
according to their scores and stored to a time-varying n-best list.
Efficient pruning on commands, words and on the acoustic level.
For
continuous word recognition, intermediate hypotheses are stored in n-best
lists ordered by their scores. If the command syntax allows it, hypothesis-
trees are grown during the recognition process, and new hypotheses are
only started if a possible word end is found. So the search space of word
and subword units can be reduced substantially and the search is only
conducted through a subset of reference models. On the acoustic level,
score based pruning reduces the number of active grid points in the
matching process [2]. All these measures reduce processing load and
allow the implementation of the recognition engine even on simple, low
performant processor platforms.
Figure 12-1.
Associative Dynamic (ASD) classifier in network representation, x primary
feature vector, y secondary feature vector, d local distance, g aggregated, optimal distance
during matching process.
Temporal compression of reference patterns.
Temporal redundancy is
avoided by compressing the reference patterns for the basic acoustic units
in a way, that the remaining reference states represent only stable and - in
terms of classification - relevant parts of the original pattern. For reasons
