Given that you can not (want not?) use extrinsic info you have to rely on intrinsic or ab-initio prediction tools eg. Augustus, EuGene, Genemark, ... and many others. The big issue here is that, in order to get somewhat good results you will have to train/optimise them for your organism, which is not a simple task (but doable though!).

I'm however wondering why you say you want to do this without extrinsic data? As jean.elbers pointed out as well, there is nonetheless valuable info in all the proteins known so far, even if they are not specifically from the you are working with. Transcript data might indeed be a little less straightforward but the protein info is gonna be for sure valuable!

Actually, you will achieve best performance of your genome annotation tools when combining both intrinsic and extrinsic info in your approach. So perhaps reconsider using any available data source.