A general architecture for a frame-based word recognition system

The main features of a multilevel frame-based word-recognition-system architecture are described. The recognition process is based on a basic phonetic-unit recognition, where the basic phonetic units contain information about miscellaneous speech features of one signal frame. The proposed architecture should make it possible to recognize a word as well as to describe its prosodic features. Here, a speech segmentation should be the domain of analytical methods rather than pattern-matching techniques. Due to the uniform structure of all recognition units and their mutual processing independency, some degree of parallel processing is possible on each processing level. Considering this, a real-time realization of the proposed architecture might be feasible.