Synthesis Filter/Decoder Structures in Speech Codecs
Jerry D. Gibson · 2012
Using the Shannon backward channel result from rate distortion theory, we derive new decoder/synthesizer structures for common linear predictive speech models. We calculate the transfer functions for these new structures for common source models and discuss the effect on the reconstructed signal. The perceptual weighting filter and the average distortion both play a prominent role. We show that common CELP encoders and decoders lack these components, and note that at least at the decoder, postfiltering as currently used in CELP may partially compensate for the missing structure.