On the relationship between root interpolation and LSP interpolation

Line spectrum pair (LSP) was first introduced by Itakura (1975) as an alternative representation for the LPC spectral information. It was later found that LSPs have very desirable quantization and interpolation properties. Interpolation of LSPs at the frame boundaries results in a synthesized speech that is smoother and has higher quality. This paper compares the two methods of interpolation: 1) interpolating the LSPs derived from the LPC coefficients vs. 2) computing the LSPs from interpolated roots of the LPC polynomial. Experimental results obtained using four to six second long male and female sentences suggest that these two sets of LSPs are effectively equivalent.