This paper proposes a new approach to signal selection in time-diversity systems. Specifically, we consider the problem of digital speech transmission over a burst-error channel using two-channel time-diversity reception. Let every speech segment (of length W) be transmitted twice so that at least one of the transmissions escapes an error burst, with a certain useful probability. Let the received speech segments be Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> and Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> . We propose an autocorrelation-maximizing signal selection procedure of the following form. Select Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> (or Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> ) as the “cleaner” speech segment according as C(Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> , W) ≧ (or <) C(Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> , W), where <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$ C(Y_u,W)=\sum_{r=2}^W({\rm sgn}\ Y_{ur}\cdot{\rm sgn}\ Y_{u(r-1)})/(W-1);u=1,2.$</tex> Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">uτ</inf> is the speech amplitude at sample τ in Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">u</inf> , W is a computational window that is typically a few milliseconds long, and sgn Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">uτ</inf> is a polarity function that is assumed to have zero mean and unit variance. The use of sgn Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">uτ</inf> instead of Y <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">uτ</inf> leads to a simply implemented selection procedure, and computer simulations have demonstrated its practical utility. For example, in one study of three-bit dpcm coding, autocorrelation-based burst-error detection proved to be more useful than a procedure where dpcm samples were error-protected on a bit-by-bit basis, rather than in blocks.