Time-domain coders use long-term predictors to exploit the quasi-periodic structure of voiced speech. However, one major drawback with classical predictors is the restriction of the delay to integer values, reducing the performance especially when the pitch period is short. This paper presents an extension of the prediction technique to encompass non-integer delays enabling more accurate representation of quasi-periodic segments and an increased flexibility in the design of long-term predictors. The use of long-term predictors with fractional delays in CELP coding produces an enhancement of the harmonic structure in high frequencies and an improvement of quality for female speakers.