media compression usually isn't perfect. For starters, data is framed with synchronization bits to recover from bad transmissions and the like.
Also, audio codecs usually use prepared huffman codebooks for the final step.
For most codecs they're defined as part of the format, few (such as vorbis) prepend the codebooks to the data stream so compressors can use different codebooks if they choose to do so.
For vorbis there used to be an experimental tool that optimized huffman coding of an existing stream, giving another 4%-or-so compression with otherwise identical data (ie. no quality change).
These 4% plus framing are probably the entropy differential they're exploiting here.
Also, audio codecs usually use prepared huffman codebooks for the final step.
For most codecs they're defined as part of the format, few (such as vorbis) prepend the codebooks to the data stream so compressors can use different codebooks if they choose to do so.
For vorbis there used to be an experimental tool that optimized huffman coding of an existing stream, giving another 4%-or-so compression with otherwise identical data (ie. no quality change).
These 4% plus framing are probably the entropy differential they're exploiting here.