Filtering of speech signals by over-embedding

Speech signals possess properties which mimic structures exhibited by deterministic dynamical systems. Furthermore they present strong non-stationarities. In this paper it is shown how a local projective noise reduction scheme, originally developed for low-dimensional deterministic chaotic signals, can be successfully applied to human speech, once an over-embedding is considered. The strong inherent non-stationarity, due to pronounced fluctuations of system parameters, is resolved as a sequence of many different dynamical regimes of moderate complexity, each one corresponding to a single phoneme. If some slowness properties in the change of the parameters are ensured, essential aspects of determinism are restored.