Computing an LDA Matrix

There are two ways of computing an LDA matrix. The one-pass version and the two-pass version. In the one-pass version, you need a lot of memory, at least 4*dim*dim*classN bytes, where dim is the dimensionality of the LDA-source-featurespace, and classM is the number of LDA classes. The two-pass version needs much less memory because it combines the accumulation data for all classes in one single accumulation matrix. Therefore it needs one additional pass over the traininig data to first compute the relative frequencies for every class.

You can use any recognizer for computing an LDA matrix, it doesn't have to match the new system's architecture, because the only thing that is done while computing the matrix, is the reading of labels and speech vectors. See the LDA making method of the LDA module for details.