We present a non-linear and non-parametric prediction algorithm for modelling the internal dynamics of the speech signal through a state-space geometric description. The proposed model extracts the intrinsic determinism embedded in the voiced speech signal from comparisons among the observed trajectories in state space. The combination of a metric criterion and frequency similarity results in a model which evolves throughout a quasi-periodic trajectory with a pitch period similar to the speech signal under analysis. Results for voiced frames of this model and comparisons with LPC-based methods are promising at medium bit rates.
展开▼