Because CSM provides a “return path" to the input data, it comes
with some strings attached. Feature extraction must be carefully thought out in
order to make the CSM analysis possible. Despite this, CSM allows for a wide range of
feature extraction methods, including many iterative approaches.
Keep in mind that by using the chain rule, (Section 2.2.4),
CSM feature extracotrs can be chained (cascaded) to produce
a much wider range of feature transformations.
In fact, one could say that any feature extraction method can be
put in a form where CSM analysis is possible. Take for example the
popular MEL-cepstrum features. These can be put in the CSM form
(See chapter 11). Chapters 4 through 11
provide a wide range of feature transformation types.
