Time-varied pitch gain predictor for tracing amplitude-varied periodicity speech

We proposed a time-varied prediction gain to achieve a better representation of speech periodicity than the traditional approach. Due to non-stationary variation of speech, the proposed long term predictor (LTP) can well trace the detailed fluctuation of speech amplitude in both transient and stationary periods. Simulation results show that the proposed first-order varied-gain LTP obtains a speech quality that is near to that obtained with the fifth-order fixed gain LTP, however, the former requires much lower computation than the later.