Updating of speech transmission parameters obtained from nonuniform analysis segments in a linear predictive speech compression system

Abstract

Listen

Translate article

A linear predictive speech compression system based on stationary autocorrelation formulation is described. This system employs pitch synchronous analysis of 4-kHz bandlimited input speech sampled at 10-kHz rate. An all-pole speech production model has been assumed and a 10th-order linear predicter is used. Quantization properties of k parameters [Markel and Gray, “On Autocorrelation Equations as Applied to Speech Analysis,” IEEE Trans. Audio Electroacoust. (April 1973)], pitch and energy from a large speech data base, and phonetically balanced (PB) words and sentences are discussed. This paper emphasizes the updating of speech transmission parameters based on relative changes in normalized energy between two consecutive analysis intervals. The transmission rate is assumed to be constant but the frame size, for which the transmission parameters are used, is time variant. Thresholds have been defined for relative changes in normalized energy and changes in analysis interval size between two consecutive analysis intervals. Before compression, parameters are extracted every pitch period for voiced speech and every 8.5 msec for unvoiced speech. However, the parameters are used for transmission purpose if and only if the above-mentioned thresholds are exceeded. The parameters are forcefully updated when frame size becomes equal to maximum frame size, a preset value, even though the threshold criteria are satisfied. The transmitted parameters are pitch synchronously interpolated at the receiver end. Results of synthesized PB sentences at different bit rates are reported.