Could make for an interesting codec idea though! A codebook based compressor that works on corpora of audio/video seems doable. More interesting would be a codec that aims not only to exploit models of human perception to throw away bits, but also exploits those models to produce "canonical" compressed forms that would be suitable for byte/block level deduplication. Of course, interesting things are hard, and errors would be hilarious. (See: Xerox and their copiers that change letters and numbers in documents...)