The problem of predictor mistracking for narrowband signals in backward adaptive ADPCM (adaptive digital pulse code modulation) speech coders is shown to arise as a result of feedback from the signal reconstruction filter to the predictor adaptation process. A class of residual-signal-driven lattice predictors (PR) is defined that guarantees tracking for all signals without regard to the order of prediction. The LR predictor enhances speech and DTMF (dual-tone multifrequency) signal transmission performance in the presence of transmission errors. Under error-free transmission conditions, a segmental SNR (signal-to-noise ratio) drop for speech of nearly 2 dB may be encountered for the LR predictor relative to the classical signal-drive lattice predictor. In most practical telecommunication applications, however, this degradation is outweighed by the improved robustness of the predictor.>