Nonlinear prediction of speech
Brent Townshend · 1991
Measurements were made of the correlation dimension of normally spoken speech from a single speaker, and the results reveal that most of the points in the state space of the signal lie very close to a manifold of a dimensionality of less than three. This result indicates that one should be able to construct a nonlinear predictor for speech that significantly outperforms linear predictors. To validate this conclusion, a nonparametric predictor was constructed which was able to produce a prediction gain approximately 3 dB better than an equivalent linear predictor. Similar improvements in signal-to-noise ratio were also observed when the nonlinear predictor was added to a simple speech coder.>