Model-based analysis of speech and audio signals for real-time processing based on time-varying lattice filters

In this paper, a time-varying analysis procedure based on lattice filters is proposed for real-time processing. The analysis procedure estimates the reflection coefficients from the signal frames successively in a way that the current frame is analyzed with respect to previous frames. For that purpose, a linear coefficient trajectory covering the current and previous frames is estimated under condition that the left-sided starting value of the trajectory is prescribed by analysis results of previous frames. The coefficients of the current frame can be determined from the trajectory at a particular point. Analyses of speech signals show that the algorithm yields good spectral modeling simultaneously with a smooth time development.