Error margins by segmentation of speech signals with autocorrelation procedures for linear prediction coding systems

Systems used to code speech signals with very low bit rates are realized as VOCODER systems. The systems require a reliable decision whether, in short term intervals, the speech signal has a pulse or a noise like characteristic. This decision allows the speech signal to be segmented into such classes. The pulse frequency is determined in the pulse like intervals. Different segmentation procedures including the effect of mixing the original signal with the inversely filtered original signal are investigated. The segmentation procedures are based on the autocorrelation of nonlinearly distorted signals. The quality of the different segmentation procedures is objectively determined by reference measurements using an optimally segmented reference pattern.