Thereby
the results of the hybrid approach have been obtained by
replacing the k-means vector quantization of the discrete
HMMs by a codebook obtained by a neural network. The
results are obtained using multiple frame input of the feature
vectors, which takes the context of the current features
into account. In average the relative error can be reduced by
about nearly 20% using hybrid HMMs instead of discrete
ones (comparing the first 2 rows of Tab.2 and Tab.3). Experiments
with continuous HMMs yield lower recognition
rates, although the number of mixtures has been optimized
again for each writer separately (Tab.2,3: CON1bmadd).
All features are combined in one single stream.