Enhancement of Linear Predictive Coding by using Residual Signal

Abstract

Speech coding is an important aspect of modern telecommunications. The primary objective is to represent the speech signal with the fewest number of bits, and a sufficient level of quality. Most of the low bit rate speech coders employ linear predictive coding (LPC), which models the short-term spectral information as an all-pole filter. The LP coefficients are obtained from standard linear prediction analysis as a function of the input samples.

The problem with LPC is that it suffers from many limitations, although it provides intelligible reproduction at low bit-rate. However, only two kinds of excitation signals are used, which gives an artificial quality to the synthetic speech. The performance is further degraded in noisy environments declaring a frame as unvoiced even though it is voiced. The aim of this thesis is to provide an enhanced version of linear prediction coding by making use of the residual signal. The research focuses on developing an algorithm that exploits the residual to reconstruct the signal. The LP analysis has been simulated; and its performance has been compared by changing the parameters (LP order, frame length, pitch estimation method). The obtained simulation results demonstrate that speech quality can be improved via exploiting the residual signal.