Sound Synthesis by Flexible Activation Function Recurrent Neural Networks

Abstract

In this paper we investigate on the use of adaptive spline neural networks, to define a new general class of physical-like sound synthesis models, based on a learning from examples strategy (in particular in this paper we study single-reed woodwind instruments). It is well known that one of the main problems in physical modeling concerns the difficulty of parameter identification and the definition of the exciter nonlinear function shape (which plays a key rule in the instrument timbre). In the proposed neural model we make use of FIR-IIR synapses followed by a Catmul-Rom spline based flexible nonlinear function whose shape can be modified by adapting its control points. This general structure can imitate an entire class of instruments by learning all the parameters (synaptic weights and spline control points) from recorded sounds. In order to obtain an efficient hardware/software implementation, the synaptic weights are constrained to be power-of-two terms while the nonlinear function can be implemented as a simple spline interpolation scheme or through a small lookup table. In order to demonstrate the effectiveness of the proposed model, experiments on a single-reed woodwind instruments have been carried out.