Cubic spline modeling of speech spectra

This paper examines the capabilities of cubic spline approximation in the representation of speech spectra as compared to the widely used linear predictive coding method. The LPC technique has proven robust in many applications, but certain difficulties remain when modeling speech signals. In particular the issues raised include the modeling of zeros in nasal sounds, model stability for sinusoidal inputs, global reaction of the model to a localized input disturbance, and interpretation of coded parameter values for purposes of editing and recognition. The cubic spline model addresses each of these problem areas. The approach used here to model each frame of speech data consists of fitting a cubic spline function to the local peaks of a log-magnitude FFT spectrum.