A Modular Neuro-Fuzzy Network for Musical Instruments Classification

A modular neuro-fuzzy network is proposed for the classification of musical instruments from the sound they produce. Each module, which is inherently a fuzzy inference system with the capability of learning fuzzy rules from data, operates on a distinct subset of input features. All sub-networks are separately initialized and trained by a two-phase strategy. First, a fuzzy clustering algorithm is applied to establish the structure of each sub-network as well as the initial values of its parameters. Then, each sub-network enters a supervised learning phase for optimal adjustment of its parameters. After learning, each sub-network encodes in its structure the knowledge learned in the form of fuzzy if-then rules. The various sub-networks are then combined in a single modular network that is able to face the complete classification task. Preliminary experimental results compare favorably with human performance on the same task and demonstrate the utility of the modular approach.