Hi, All
This is the mesg I received from my machine learning teacher
answering my question of nnet vs hmm. He is not an expert on
speech recognition but a general academic in the field of machine
learning as far as I know.
Just some opinions from others. (I took his name away so that
I do not have to get his permission ...)
-- Jinping
quote begins here
---------------------
Jim,
When we say that ANNs are 'very capable', we mean that they are able
to represent, or implement, any fucntion (with a sufficient number of
hidden units). This does not guarantee that they will generalize
correctly from training examples. In fact, quite the opposite is
true: as the representation power grows, the bias diminishes, and the
ability to generalize goes down. If you reduce the representation
power (increase the bias) of an ANN by reducing the number of hidden
units, the network will perform some generalization, although it is
not guaranteed to be the 'correct' generalization.
To do voice or speech recognition well, we need learners that
generalize correctly. It is an empirical question which of ANNS and
HMMs generalize better. But HMMs have a huge advantage in that they
are very easy to train, much easier than ANNs, especially for large
amounts of data. For this and other reasons, they are generally
preferred, although ANNs are still used for various speech related
tasks. |
