Comparison of symbolic and connectionist approaches to eliminate coarticulation effects in phonemic speech recognition

Two methods to correct phonemic transcriptions produced by the acoustic processor of a speech recognition system are described and compared. The first method that was invented by Prof. Teuvo Kohonen and named the Dynamically Expanding Context (DEC), involves a large set of error-correcting rules automatically constructed from examples. This symbolic approach is compared with a connectionist one, which employs a multi-layered feed-forward network trained with back propagation. Our experiments demonstrate that the latter paradigm is far from optimal when context-dependent mapping (e.g. correction) from one set of symbol strings to another is desired. The DEC-method is shown to have better correction capabilities. Beside this, the trainingtime required by DEC is a fraction of that required by back propagation.