Skip to Main Content
This paper presents a study of an LPC analysis procedure for speech which selects the parameters of an acoustic tube model--the number of tubes, the areas, and the tube lengths--by using a local fidelity measure. PARCOR coefficients are generated using Burg's method for a 10 tube model. Successively 3, 4, 5, 6, or 7 of them are forced to zero or a constant value, and the combination with the least distortion is selected for each case. From among these, a final selection for each frame is made by rule. Other issues considered include the effects of different fidelity measures, quantization, and non-exhaustive search schemes.