Both very popular differential expression packages assume a raw gene matrix count. This makes sense because the statistical model models the library depth. But what if I normalize my counts with say, the ERCC RNA spiked-in standard? Would that affect the statistical power? False-positive? Why exactly is it not recommended to give a normalized count matrix?
The advice is to not generally use ERCC spike-ins at all because of variations introduced by pipetting at the volumes they recommend.
The thread also explains how to use DESeq and EdgeR with spike-in normalisation, with the process being easier significantly with DESeq, where you can use the calcSizeFactors on a count matrix of spike-in reads alone. With edgeR you will have to pass values using the lib.sizes parameter in the apposite functions.
If you want to use fractional counts through limma-voom should work, I think; I've had good results using voom on RSEM counts, which are fractional.