This paper proposes a new connectionist model of supervised learning called WYLINWYT map, based on Kohonen's Self Organizing Map (SOM). SOM has been early used in speech processing. It produces, by an unsupervised learning, a topological representation of the speech input space. The unsupervised learning mode is not fitted to obtain acceptable results in speech recognition. WYLINWYT shows a new learning scheme that improve speech recognition capabilities of SOM. WYLINWYT combines the advantages of a topological representation in the map space and the discriminating power of supervised learning. Phoneme recognition experiments are performed in a corpus of 200 phonetically balanced sentences. Compared to basic SOM, the proposed recognition method shows an improvement in the recognition accuracy of more than 5%.