We have studied experimentally the operational rate-distortion performance
for very low bit-rate speech coding using linear inter-frame dependencies.
We propose an algorithm that efficiently combines quantization and linear i
nterpolation procedures. With a maximum delay of 200 ms, for the spectral e
nvelope information and using line spectrum pair (LSP) parameters as input
space the proposed algorithm performs best at rates of between 200 and 300
b/s. For comparison's sake several other procedures such as the multi-frame
encoder (Kemp D., Collura J., Tremain T., Multi-Frame Coding of LPC Parame
ters at 600-800 bps. In: IEEE ICASSP-91, 1991, pp. 609-612) and matrix quan
tizer (Tsao C., Gray R., Matrix quantizer design for LPC speech using the g
eneralized Lloyd algorithm. IEEE Transactions on Acoustics, Speech, and Sig
nal Processing ASSP-33, 1985, 537-545) are simulated. Furthermore, a mono-d
imensional version of the proposed procedure is shown experimentally to pro
vide the best operational rate-distortion trade-off when coding a parametri
c representation (pitch, gain and voicing information) of the excitation si
gnal. (C) 2001 Elsevier Science B.V. All rights reserved.