A Differentiable Acoustic Guitar Model for String-Specific Polyphonic Synthesis

We introduce a differentiable model for acoustic guitar synthesis. The model takes in a MIDI-like conditioning representation, in which the guitar string for each note is specified, and synthesizes polyphonic audio. We employ a Differentiable Digital Signal Processing (DDSP) -style monophonic synthesis module, in which parameters are predicted that drive inharmonic oscillator and filtered noise synthesizer modules. Our monophonic synth is conditioned on a guitar string index, so 6 individual string audio signals are produced, summed, and fed through a trainable reverb to produce the final polyphonic audio. We train our synthesizer using data from the GuitarSet dataset and observe that the output audio is expressive and reflects the timbre and recording environment of the target instrument from the dataset. We conduct a listening test and evaluate the synthesizer’s ability to reconstruct unseen audio, comparing against an off-the-shelf acoustic guitar synthesizer, and a sample-bank synthesizer constructed from excerpts from GuitarSet. Additionally, we demonstrate how synthesized audio can be modified to vary parameters from the performance, including simulating use of a capo and alternate fretboard positioning.