Does anyone know, or can give an approximation, as to what the actual gain from the lossless coding step in modern codecs is?

In other words, if you'd throw out the huffman coding and replace it by a simple bitpacker, what kind of efficiency loss would you suffer?

Scott Hacker mentions 20% in his book. (MP3: the definitive guide)

On the codec I implemented as thesis the average gain was 15% with arithmetic coding and a uniform model.

Of course that a well tunned model should increase this gain, but I'm not sure by how much, probably somewhere between 5 and 10%.

IMHO cca. 20% too, for cca. 130 kbps vorbis.

20% seems to have been correct.

Is there any formal literature / published research on the subject?

I think that this could be tested (and be interesting) by modifying a decoder so it computes and reports the ratio before/after Huffman decoding.

Btw Scott Hackers' 20% mention is probably not far from the truth, but he might have written 20 because I told him that it SHOULD be 20 on average.

I already did this Gabriel , hence the "seems to have been correct" reply

I am wondering if there is literature or published research about this, or alternative approaches. Of course there is general literature, but I mean audio coding specific.

Mind you disclose the tool?
It could be interesting to compare Huffman efficiency of different encoders.

Regarding litterature, I am not aware af any regarding this. It was probably discussed in the 90's when standardizing MPEG-1, but info from this period is quite scarce.

Can't do that. I'll clarify in PM.