Enhanced spectral modeling for MBE speech coders

Low rate speech coders based on MBE or STC models require an accurate representation of a discrete set of significant spectral amplitudes for proper performance. Although compact and easy to code, classical linear predictive spectrum representations model adequately only the general short time spectral envelope, rather than the desired discrete amplitudes. A technique for objectively enhancing the accuracy of a linear predictive spectrum representation through the use of cubic spline interpolation and amplitude compression is described, along with a perceptual enhancement technique based on mel-warping the spectrum. When incorporated into a 2400 bps MBE-type coder (EMBE 2400), the objective enhancements applied to an order-14 LP model are shown to be superior, in terms of spectral distortion, to an unmodified order-22 LP model. Additionally, the perceptual enhancement is shown to provide reduced spectral distortion in the relevant frequency band.