Just some ideas:

If your data is not ribo-depleted, first filter against an rRNA database, e.g. sortmeRNA.

If you have a reference genome and annotation: map against that, remove all hits against known protein-coding RNA, assemble the remaining reads.

If you don't have a reference genome/annotation, assemble the reads, remove all transcripts with ORFs that yield good matches against protein-coding RNAs, e.g. by blastx against nr.

For identification you could run Infernal on the set of remaining transcripts.