In a paper submitted to arXiv, Pablo Tablas de Paula and colleagues present a self-supervised, event-based sound matching model built around a differentiable extended Karplus-Strong algorithm. The aim is to recover interpretable synthesis parameters from audio, particularly plucked guitar, without relying on external onset and fundamental-frequency detectors. The authors explore training strategies that combine parameter losses on synthetic data with audio losses.
They show that time-domain fractional delay interpolation provides gradient accuracy comparable to frequency-sampling while avoiding time-aliasing in highly resonant, time-varying scenarios. Systematic gradient analysis reveals that standard spectral losses provide no meaningful directional gradients for onset times, which heavily degrades joint training. Training exclusively with parameter losses on synthetic data effectively learns fundamental frequency, timbral parameters, and onset times, but struggles to generalise to monophonic studio recordings of plucked guitar.
The best generalisation comes from combining external detectors with audio losses, which isolates the model to timbre optimisation. The Karplus-Strong decoder recovers interpretable parameters and naturally captures the transient characteristics of plucked guitar, but Harmonics-plus-Noise baselines yield higher reconstruction fidelity by most metrics. This highlights a tradeoff between interpretable synthesis and raw reconstruction quality.