The Burg algorithm for computing reflection coefficients of speech (sometimes called the covariance lattice) is compared with the autocorrelation method of linear prediction. Certain problems for small window sizes for both methods are discussed. Three dimensional displays of spectral representations and audio tapes of synthesis are used to illustrate the practical result that there appears to be no advantage to either approach for speech processing, other than the computational efficiency of the autocorrelation method. [Work supported by the Department of Defense.]