Aliasing-Aware RTNeural-Compatible WaveNet Modeling of Guitar Amplifier and Pedal Captures

This technical preprint reports an applied study of neural guitar amplifier and pedal modeling under a strict real-time export constraint. Trained models must be representable as RTNeural JSON, validate against TensorFlow/Keras reference inference, and run as causal low-latency DSP in a native C++ audio runtime. The study compares dense, recurrent, and convolutional candidates, and reports that WaveNet-style temporal convolutional networks produced the strongest end-to-end evidence in the tested 48 kHz mono paired-capture dataset. The paper evaluates clean, edge-of-breakup, crunch, rhythm, lead, and overdrive-pedal captures using waveform error, residual statistics, native real-time factor, and probe-based aliasing-to-signal ratio. The work is released as an open technical report associated with the RTNeural Trainer project.