I was a little surprised at how badly I did at this, given that there was almost no delay between reading and reciting each sentence.
The errors I make are interesting:
- Word substitutions that don't change the meaning of the sentence.
- Insertion of common words that don't change the meaning of the sentence.
- Rearrangements of phrases in the sentence. Sometimes this changes the meaning, though the result is always grammatical.
- If there are several similar elements (eg numbers, lists of items) they may be shuffled, and some may be dropped.
- Sometimes a phrase is repeated. (eg something that works as both a prefix and suffix gets said twice)
Some sentences are much harder than others to regurgitate, even among sentences of the same length.
I have no doubt that this is a skill that can be learned.
The errors above are a demonstration that memory recall is an act of resynthesis, a creative act.
Therefore, practicing memory tasks such as this one should increase performance on other related creative tasks. In this case, assembling eloquent sentences in real-time.
At the very least, it's giving a larger pool of examples to resynthesize from. I don't think you get the same effect from reading: you don't need to retain the exact phrasing to retain the meaning, so specifics of phrasing can be discarded early in processing. Reading while looking at the page also will not work, as it does not require that act of creative recall. Read-and-recite makes sure the sample data gets run through the whole pipeline.
A similar visual exercise would be to copy an illustration in front of you, copying it in as large chunks as possible. This should help you draw creatively. I expect the errors would be very reminiscent of resynthesis.