Recently, I have tried to apply GOSeq to RNA-seq expressions after voom adjustment. It has just ocurred that differential expression and gene length were negatively correlated, what is unexpected. I was wondering how much of the gene length impact on Differential Expression Analysis is already removed when using voom.  Can someone point me to literature regarding this issue ? 

There isn't any literature on this because there is nothing to say. voom does not remove gene length bias in DE results.

None of the good RNA-seq DE methods remove length bias, and it would not be at all desirable for them to try to do so.

If your significant genes tend to be short genes rather than long genes, then that is a property of your data. It has not been caused by using voom. Any other good DE method would likely give a similar trend.

I personally would not do a goseq analysis if the length bias trend looks to be decreasing. If I saw that, then I would turn off the goseq bias correction, or else go back to my DE analysis to see if there is something unexpected about the data that I had overlooked.