Implicit neural differentiable FM synthesizer
The purpose of this project is to emulate arbitrary sounds with FM synthesis, where the parameters of the FM synth are learned by optimization.
This idea was conceived and implemented during the Neural Audio Synthesis Hackathon 2021. Thanks to Ben Hayes for organizing the workshop and to Mia Chiquier for pointing me towards SIREN!
Architecture
Please refer to FMNet and Envelope in synth.py for the actual architectural details.
This model takes as input a list of time steps t_1, t_2, ..., sampled at some sample rate, and outputs an audio signal in the same sample rate.
Similar to SIREN, it feeds the input time step values through sinusoidal activation