A novel method to explore the sparse nature of prediction parameters

Compressed sensing is a new paradigm to explore the sparse nature of the signals. Compressed sensing allows to acquire signals fundamentally below the uniform rate digitization followed by compression usually used for storage and transmission. Linear Predictive Coding is the core of most of the speech coding algorithms. In this paper a novel method is presented to explore the sparse nature of the predictive parameters. Compressed sensing is applied on to the prediction parameters. Parameter estimation is posed as a 0-norm minimization problem. Alternative representations of the linear predictive parameters, i.e. the reflection coefficients and the line spectral frequencies are also considered and the performance is analysed using the spectral distortion measurement. The number of reflection coefficients and line spectral frequencies used for encoding can be reduced using compressed sensing.