Application of extremum sampling in speech coding

The magnitude spectrum of speech is sampled to extract the extremum (maximum) points representing the underlying sine-wave amplitudes. The resulting nonuniform samples are, first, interpolated using a cubic spline function and, then, modelled by an all-pole magnitude spectrum. The gain factor and the line spectral frequency (LSF) domain representation of the coefficients of the all-pole model are quantized at the encoder. A faithful reconstruction of the spectral extremum envelope is obtained at the decoder using the dequantized all-pole model parameters