Spectral templates might be adjusted by multiplying them by a vector of factors, and cepstral templates by adding a vector of terms. However for maximum flexibility the implementation should allow arbitrary adjustment. One way to do this would be to have the recognizer call a function on each template (model, reference vector, etc..) before using it for recognition. This function would be implemented within the preceding "scene analysis" stage, rather than within the recognizer. By default it would do nothing but this behavior could be overridden.
This implies of course a two-way exchange of data between the recognizer and scene analysis modules, which is of course troublesome.