Truly novel transcripts are not part of the annotation, so you haven't counted or encountered any with your htseq-deseq2 pipeline. You could use Trinity in genome-guided assembly mode. Then you could compare the generated transcripts with annotated transcripts, filter for near-identical overlap with existing transcripts, and also annotate them for their coding potential, e.g. with transdecoder+trinotate. Those that come out without a "good" coding region and e.g. no blastx similarity to known proteins in GenBank or without InterPro domains, could be candidates for a list of a) novel transcripts and b) non-coding transcripts.

Which genome is this by the way? For well-annotated model organisms that approach might not yield much.