Usually, the appraisal of antique violins such as Stradivarius and Del Gesu are conducted by detailed inspection of the shape and workmanship. On the other hand, some violin connoisseurs and violinists have certain impressions of Stradivari's timbre, and they believe they can distinguish between a Stradivari and others. Therefore, we tested the possibility of sound-based appraisal. The performance sounds of 13 violins including four Stradivari violins and new ones were recorded in a studio and the acoustic features (spectral envelope and mel-frequency cepstral coefficients, MFCC) were used to train a neural network. A variational autoencoder (VAE) was used for training on the acoustic features and clustering of violins in the latent space. Visual identification of the violins was conducted by drawing a set of points in the latent space decoded by the VAE. The results show that the MFCC produces the best identification accuracy. Instruments by the same maker were distributed close together in the latent space. Nevertheless, there were recording conditions under which the VAE could not generate accurate identification.
Yokoyama et al. (Wed,) studied this question.