Music Rhythm Recognition Through Feature Extraction and Neural Networks

In this paper a procedure to solve the problem of recognition and classification of sampled musical rythms is presented. The lack of precise rules for doing this analysis makes difficult and often ambiguous the automatic execution of a cognitive process naturally performed by human brain. This procedure can be extended to the classification of any signals showing similar characteristic (i.e. EEG or ECG). Due to the complexity of the time dependence, standard procedures used for chaos characterisation (i.e. correlation dimension, Lyapunov exponents, etc) can fail. Moreover a direct usage of artificial neural network can introduce too many optimization variables. The proposed procedure can be organized in two phases : the extraction of some new type of invariant from the sampled time series and the usage of this extracted features as input for a classifying standard neural network. This system was able to distinguish between binary and ternary signals with a precision of 99 %. The single rhythm was classified within an error of 5 %. This system seems to be able to deal with the behaviour that characterises a musical rhythmic sequence, and to classify patterns independently of the musical instrument and tempo.