Using LDA (Linear Discriminant Analaysis)

This page is not meant to explain how LDA works. Using LDA means doing a matrix multiplication on each feature vector as a preprocessing step. LDA is treated separately from the other preprocessing steps because LDA needs the entire training set's labeled speech data. There are different ways of computing an LDA matrix. If you want to use LDA you first have to compute the matrix before you can start a regular training. For a very initial system it is not necessary to use LDA, in fact, you better not do it to avoid too many sources of possible errors. Once you have a nice system, you can still start over with an LDA in the front end, that was computed on reasonably good labels.