| The problem of transmitting line spectral frequencies (LSF) generated by the Federal Standard 1016 CELP speech encoder over a packet-loss network is considered. Multiple description (MD) coding techniques are investigated, which reduce the spectral distortion (SD) in speech due to packet losses while using the same transmission rate as the standard CELP encoder. We focus on exploiting the residual redundancy of the encoder output to estimate lost packets at the receiver, by using hidden Markov modeling of the encoder output and estimation based on the forward-backward algorithm. In particular, the problem of optimizing index assignments for Markov decoders is addressed. Experimental results are presented which compare the proposed techniques with other known techniques, such as linear estimation and Gaussian mixture modeling. It is demonstrated that the proposed Markov technique averages 2.18 dB of SD when one description is lost, compared to interpolation of odd-even split LSFs at 2.99 dB. |
