AUTOMATIC-INDEXING OF A SOUND DATABASE USING SELF-ORGANIZING NEURAL NETS

This article concerns the organization of a set of sounds without the use of predefined attributes. The approach is based on a neural network called Kahonen feature map (KFM). Teuvo Kohonen has developed an algorithm for this network, which allows the mapping of an input space onto a topology-preserving feature map in an unsupervised, self-organizing learning process. The KFM algorithm was extended to process dynamic sounds which can be use as a retrieval index for a sound archive or to control a synthesizer. The sound feature map (SFM) is demonstrated by a small sound database system and a simple feature map synthesizer control system.