LPC-Spectral convolutional variational autoencoder for violin sound mechanism analysis
Abstract
As one of the most expressive string instruments, the violin's complex sound-production mechanism and rich timbre variations have posed research challenges in the field of musical instrument acoustic modeling. Existing physical modeling methods struggle to capture the nonlinear characteristics of bow-string interactions. To this end, a digital modeling method for violin sound mechanisms was proposed based on LPC-Spectral Convolution-VAE. First, a multidimensional acoustic feature extraction method was designed that fuses LPC and spectral convolution, capturing formant structure and vocal tract characteristics via LPC analysis, combined with multi-scale spectral convolutional networks to learn harmonic distributions and spectral textures. Second, a VAE network-based bow-string interaction dynamic modeling mechanism was proposed to establish probabilistic mapping from performance control to acoustic features through variational inference, achieving natural and diverse timbre generation. Finally, an LPC-Spectral Convolution-VAE modeling framework was constructed to implement a complete process from feature extraction and interaction modeling to audio reconstruction. Experiments on the URMP and ViolinSet datasets show that our method outperforms existing methods in spectral distortion, timbre similarity, and performance expressiveness metrics, reducing spectral distortion by 13.6% compared to the best baseline method to 4.78±0.54 dB and improving timbre similarity to 0.921±0.026, thereby generating realistic violin timbres.