FAQ

FAQ

Quick answers. Anything install-related lives in Installation; failures live in Troubleshooting.

Does the ASR need a GPU?

No. The local compose stack runs faster-whisper on CPU (ASR_WHISPER_DEVICE=cpu, int8 compute). services/asr_service/Dockerfile.gpu exists for CUDA deployments (e.g. Cloud Run with an L4 GPU): it sets ASR_WHISPER_DEVICE=cuda. CPU transcription is slower but fully functional.

Can I run AIrport without X-Plane?

Yes, partially. The whole voice → readback loop works without the simulator: the HMI chat shows the transcription and the pilot's reply, flight strips and clearances update. What you lose is aircraft motion on the 3D field and the spoken TTS readback — both happen inside X-Plane via the plugin.

Can I run it without the Cloud Run agents?

The stack starts fine with empty *_AGENT_URLs, but dispatches produce no pilot replies — the agents are the pilots. Deploy them once (guide); they are stateless and scale to zero, so an idle deployment costs almost nothing.

Which airports are supported?

• LEST (Santiago) is the default airport in the HMI and the arrival simulator.
• LEBL (Barcelona) ships pre-fetched in data/airport_data/ (LEBL.dat, LEBL_graph.json).

Airport data (ground network, stands) is downloaded on demand from X-Plane's gateway via xplane-airports, so other ICAO codes can work — but taxi routing and stand assignment are only validated on these two. Adding an airport means verifying its ground graph; start at Shared → taxi_router.

Where are the models and datasets?

• ASR: jacktol/whisper-medium.en-fine-tuned-for-ATC-faster-whisper (downloaded on first run, ~1.5 GB).
• Readback corpus: santiisoutoo/pilot-readback-corpus on Hugging Face (audio version available on request — see the README).
• Evaluation code: agents_evaluation/ — see Data & Testing.

License and citation

Free for non-commercial use — see LICENSE. If you use AIrport in research, cite the paper; BibTeX in the README.

Related

Installation · System Overview · Quickstart