Predictive fractal interpolation mapping: differential speech coding at low bit rates

Abstract

In traditional linear prediction coding of speech, the vocal tract is represented by an AR model and the model is based on sample prediction. This paper proposes a new prediction model using iterated function systems (IFS) for vector modeling of arbitrary discrete sequence. The inverse algorithm for model parameter estimation is addressed, which is different from the inverse problem of the normal linear fractal interpolation. Furthermore, a new differential vector coding scheme using the new technique is presented, which also broadens the IFS implementation because of the backward prediction and differential vector usage. The model of predictive fractal mapping has very good potentials for low bit rate speech compression, for instance, vector prediction mode and less model parameters related to the waveform shapes. It is a new area of speech coding, particularly, at low bit rates. The coding scheme presented is novel and unique and has great potential applications.