Fixed (nonadaptive) and forward adaptive differential pulse code modulation of regenerative composite sources is investigated. In the fixed code, an approximate formula is given for the optimal value of the prediction coefficient. This is then used as an initial guess to optimize the code (predictor and quantizer) through a numerical method. In the forward adaptive scheme, the state of the switch in the composite source is estimated using a MAP sequence estimation algorithm, and the code is then matched to the mode process corresponding to the estimated switch state. The performance of the two systems is evaluated with quantizers of 4, 8, and 16 levels. The results show that the forward adaptive scheme significantly outperforms optimized fixed DPCM in the sense of mean-squared error. Stochastic stability of the code is also established for the fixed DPCM scheme as well as for an adaptive scheme which receives the switch state as side information.>