Application of polynomial splines to a time-varying autoregressive model of speech

An investigation is conducted into the use of polynomial spines in a time-varying autoregressive model of speech. The form of the model and computational considerations are discussed. Results are presented for the analysis of synthetic signals and real speech. The effects of model parameters, in particular, spline order and knot spacing, are discussed. The spectral evolution as provided by this model is compared with the more conventional stationary models. The performance of this model for tracking time-varying autoregressive coefficients is compared with the models which use an orthogonal time-series expansion of autoregressive coefficients.<
>