I've done this by doing a RPkM-like transformation which I call VSTPk. Transformations that adjust for transcript length will correct a little bit for the noise associated with mapping short reads to transcripts of varying lengths:
Another, probably better, alternative is to use Salmon to map reads to transcripts, which applies its own mapping-based correction to read counts. Salmon produces a statistic that relates to the mappable length of a transcript, and normalises counts based on how many molecules the model predicts came from each transcript isoform.
Unfortunately all these normalisation techniques require good gene models, which are difficult to obtain for non-model organisms.