Multi-Level Multi-Decision Models for Automatic Speech Recognition is discussed. 
It is hierarchically organized. Here there are not used the generative grammars 
for model speech signal synthesis as a feedback in speech recognition process. 
Instead of the latter multiple decisions, but under simplified conditions, 
at all levels of a speech signal processing hierarchy is introduced. The 3-level model 
with phoneme recognizer, word recognizer and continuous speech interpreter is proposed. 
Experimental results for the 3-level model are given and problems to be solved are discussed.