You could also try MLAT with the ATC comms to get a plausibly more deterministic source for the transmissions. I've done a decent amount of work with voice recognition for fire/police and the technology just isn't quite there yet. ATC comms probably have a lot less variation than police/fire calls, so you could probably train some custom models with a bunch of labeled ATC data to get you in the vicinity.

Couple other sources:

- Passive radar - Some pretty interesting results in this area https://www.rtl-sdr.com/tag/passive-radar/

- Sound detection/recognition/localization - Could correlate with ADSB to label sounds with known airframes to help train an identification model

- Video detection