Musical visualization and F0 estimation using neural network

This paper investigates how to extend ability of feed forward neural network for purposes of musical note visualization and F0 estimation. We set experiments to find the best features that introduce high generalization rate and good F0 estimation result per single audio frame. These features were extracted from spectral data, autocorrelation, auditory filter bank, and modified Ceptral methods. The samples in our experiments were generated using real musical instrument sound recordings. To compare all investigated features, we trained 56 neural networks with random mixtures up to 4 simultaneous notes and evaluated with both random note combinations and chord patterns. The experiments shown that using features from auditory filter bank, our system gives better estimation results for musical instrument signals with variations in both amplitude and phase. Finally, we evaluated visualization of our system using audio signals from both synthesizer and CD recordings.