In this contribution, a time-varying linear prediction is proposed for speech analysis and synthesis. In comparison to the time-invariant prediction, the predictor coefficients are time-varying within the frames. For that purpose, the coefficient trajectories can be described by basis functions. This approach leads to discontinuities between the frames if the frames are analyzed independently. Therefore, continuous conditions are defined which force continuous trajectories also between the frames. The estimation of the optimum coefficients of the basis functions is solved analytically by a least mean square approach. The analysis results show that the estimation algorithm achieves smooth trajectories of the vocal-tract resonances together with a high time resolution, which is interesting for a variety of application.