LPC spectral moments for clustering acoustic transients

Spectral moments (mean and coefficients of variation, skewness, and kurtosis) are assessed for 40 samples from 10 groups of acoustic transient signals differing in harmonic structure, duration, and degree of spectral overlap. Discriminant analysis involving moments based on linear predictive coding (LPC) resulted in a higher recognition rate for pulsed-tone sounds (87%) that were more like human speech than for pure-tone sounds (70%). By contrast, classification based on moments calculated from the discrete Fourier transform (DFT) yielded 85% recognition for both groups. Cluster analyses indicated that LPC-based moments were more characteristic of relationships among the 10 sound groups and especially the two tonal groups, though results were somewhat dependent on LPC model order.<
>