A nonstationary signal source can be decomposed into subsources if it exhibits certain characteristics. Adaptive predictive coding is considered as a method of extracting these subsources from the composite source in such a way that the overall communication problem can then be viewed as two different, but connected, communication problems: 1) transmission of predictor parameters, and 2) transmission of the subsource using difference signals. An intermediate fidelity criterion is defined that describes the effects of predictor parameter distortion on the reconstruction of the difference signal. Then a rate-distortion bound on the channel requirements for the transmission of the predictor parameters is found, subject to a dual fidelity criterion. The approach is applicable to a wide range of nonstationary signal classes. The speech process is chosen here because of the wide interests in the information content of speech. The signal class consists specifically of selected voiced phrases.