If the generative model can accurately generate sounds matching a specific underrepresented specie, it must have somewhere an embedding that does separate well

If the generative model can accurately generate sounds matching a specific underrepresented specie, it must have somewhere an embedding that does separate well and could be used for classification by clustering. No?