This paper is intended to give an outline of the development and the state of the art in acoustical analysis of speech for recognition purposes. Starting with a short view on the pattern recognition problem in general, the paper shows how the spectral primary analysis will be more and more refined by including knowledge on the speech production and perception processes. Recognizers with high performance need a secondary analysis which produces feature vectors with components which are phonetically or articulatorically meaningful. Because it will be complicated to elaborate a more or less complete set of detectors, universal procedures are required for producing this description. Several approaches are discussed.