A combined time-varying and time-invariant prediction algorithm based on lattice filters for speech analysis and synthesis

In this paper, an estimation procedure is proposed which yields an optimum solution of a combined time-varying and time-invariant prediction in a non-iterative way. The prediction is performed in terms of reflection coefficients of a lattice filter. The estimation of the time-varying trajectory of reflection coefficients is based on an expansion of the trajectory by basis functions. The time-varying trajectory is assumed to be piece-wise linear. The coefficients of the time-varying basis functions are coupled with those of time- invariant estimations. The combined time-varying and time-invariant analysis is solved by a least squares estimation. The analysis and synthesis of speech signals show that the combination of the time-varying and time-invariant analysis is advantageous since a smooth trajectory together with a high time resolution can be achieved.