Rate distortion performance bounds for wideband speech

We develop new rate distortion bounds for wideband speech sources based on phonetically-motivated composite source models, conditional rate distortion theory, and perceptual wideband PESQ (WPESQ) distortion measures. The approach is to calculate rate distortion bounds for MSE distortion for each subsource of the composite source model and use conditional rate distortion theory to calculate the MSE R(D) for the composite source. Since MSE is not a useful distortion measure for today's best-performing voice codecs, we generate a mapping of MSE-to-WPESQ using fully backward adaptive waveform coders, which have MSE distortion values that correctly order their performance, and for which WPESQ values can be generated. We generate the final rate distortion functions with the mapping and show that our new rate distortion curves lower bound the performance of the best known standardized wideband speech codecs.