Contribution
Neural Modeling of Control Signal Contours in Wind Instrument Synthesis
* Presenting author
Abstract:
In sound synthesis, control signals describe time-varying parameters such as fundamental frequency, brightness, and gain that shape the expressive character of a performance. When these signals evolve smoothly over time, they form control signal contours, which are essential for producing natural-sounding wind instrument performances. Abrupt changes and unnatural transitions between notes in these contours often result in unrealistic articulation and audible artifacts. We propose an autoregressive neural model that predicts smooth, frame-wise control signal contours from sequences of discrete note events. Given note-level attributes such as pitch, onset, duration, and velocity, the model estimates trajectories for fundamental frequency, brightness (e.g., spectral centroid), and gain. These contours serve as input parameters for pulsetable synthesis, a waveform-based approach that produces realistic timbres by modulating pulse-based signals according to the predicted control trajectories. The neural contour model is trained on paired datasets of MIDI-like note events and aligned frame-wise control signals extracted from multi-track recordings of wind instrument performances. To improve generalization, we apply musically meaningful data augmentation covering variations in tempo and key. Initial results show that our approach reproduces expressive phrasing and articulation patterns characteristic of real performances.