I'd generally align to the transcriptome using bowtie2 for each sample, run picard tools (CollectInsertSizeMetrics.jar) on the resulting file. Get the number for
MEAN_INSERT_SIZE, then work out
MEAN_INSERT_SIZE - (2 x read length)
I'd generally just run BBMap, which does not require that parameter, as it calculates it automatically. Of course, I'm somewhat biased, as I wrote BBMap. But I have yet to find a scenario where Tophat is a desirable program to run over the alternatives (BBMap or STAR).