Data compression of Linear Prediction (LP) analyzed speech by interpolation and quantization of pseudo-formant parameters
William J. Sanders, Carolyn Gramlich, Arvin Levine, Robert Laddaga, Patrick Suppes · The Journal of the Acoustical Society of America · 1979
Pseudoformants and bandwidths, derived from LP coefficients, prove highly susceptible to data compression techniques such as interpolation and quantization because of their low entropy and smooth sequential nature. We have developed a heuristic to label initially unordered pseudoformants in a way that reveals their sequential smoothness, so that linear interpolation can be successfully applied to each pseudoformant separately. A linear least-squares fit is made over several points; the number of points is limited by a bound on the allowed mean-squared fractional difference between the data and the interpolating line, and another bound on the maximum-squared fractional difference. Quantization of the resulting parameters is constrained not to introduce more error than was allowed in the interpolation procedure. All parameters need not be interpolated using the same bound; if the relative importance of the parameters is chosen in an acoustically reasonable way, interpolation can produce very high quality sound at considerably reduced bit rates. [Work supported by NSF.]