Linear Prediction and Long Term Predictor Analysis and Synthesis

Abstract

Spectral analysis may not provide an accurate description of speech articulation. This article presents an experimental setup of representing speech waveform directly in terms of timevarying parameters. It is related to the transfer function of the vocal tract. Linear Prediction, Long Term Predictor Analysis, and Synthesis filters are designed and implemented, as well as the theory behind introduced. The workflows of the filters are explained by detailed and codes of those filters. Original waveform files are framed with Hamming window and for each frames the filters are applied, and the reconstructed speeches are compared to original waveforms. The results come out that LP and LTP analysis can be used in DSPs due to its periodical characteristic, but some distortion might be coursed, which examined in the experiments.