twangodev/tartanaviation-atc-labels

TartanAviation ATC ASR Labels

Machine transcripts and confidence scores for twangodev/tartanaviation-atc-adsb-utterances. A 1:1 labels-only add-on (no audio): one row per source utterance, same shards and row order, keyed by utterance_id.

531,050 labels · 184 shards · 100% coverage · ensemble ASR + weighted ROVER + ADS-B callsign snap · ~326 human-reviewed. Built with readback.

Usage

Join 1:1 onto the source. Rows are aligned and in the same order:

from datasets import load_dataset, concatenate_datasets

src = load_dataset("twangodev/tartanaviation-atc-adsb-utterances", split="train")
lab = load_dataset("twangodev/tartanaviation-atc-labels", split="train")

joined = concatenate_datasets([src, lab], axis=1)   # or join on utterance_id
joined[0]["audio"], joined[0]["transcript"], joined[0]["confidence"]

Schema

column	type	notes
utterance_id	string	join key; matches the source 1:1
transcript	string	fused transcript; "" for non-speech; human edit when review_status=edited
confidence	float	quality score in 0 to 1 (ranking, not calibrated; see below)
review_status	string	auto / verified / edited / rejected / non_speech
agreement_score	float	inter-model agreement, 0 to 1
rover_confidence	float	ROVER mean winning-vote fraction, 0 to 1
advisory_disagree	float / null	disagreement of an out-of-vote model with the fused text
n_models_agree	int	size of the agreeing plurality
callsign_matched / callsign_tail	bool / string	whether an ADS-B callsign was snapped in, and which tail

review_status: auto (ensemble only, the vast majority) · verified (human accepted) · edited (human corrected) · rejected (human flagged unusable) · non_speech (silence/noise, transcript=""). Human verdicts set confidence to 1.0 (0.0 for rejected).

How labels were made

Three ASR models, parakeet-tdt-0.6b-v2, canary-qwen-2.5b, and whisper-large-v3-atco2-asr, fused with weighted ROVER, callsigns snapped from ADS-B, and ~326 clips human-reviewed. confidence is the mean of agreement_score, rover_confidence, and 1 - advisory_disagree. These are machine labels, not gold ground truth (except verified/edited).

Confidence

confidence ranks labels reliably but is not a calibrated probability. It was tuned on a worklist of deliberately hard or ambiguous review clips, not a random sample. On that (hard) slice the human accept-rate was ~91% at >=0.8 and ~97% at >=0.9, so true rates on the full corpus are higher. Pick your own cutoff, or use the raw signals and review_status to build your own filter.

License

CC-BY-4.0, inherited from CMU AirLab's TartanAviation. No audio is redistributed, only derived text labels. Labels generated by nvidia/parakeet-tdt-0.6b-v2 (CC-BY-4.0), nvidia/canary-qwen-2.5b (CC-BY-4.0), and jlvdoorn/whisper-large-v3-atco2-asr (Apache-2.0); model outputs are unrestricted. Credit CMU AirLab and cite:

@article{patrikar2024tartanaviation,
  title={TartanAviation: Image, Speech, and ADS-B Trajectory Datasets for Terminal Airspace Operations},
  author={Jay Patrikar and Joao Dantas and Brady Moon and Milad Hamidi and Sourish Ghosh and Nikhil Keetha and Ian Higgins and Atharva Chandak and Takashi Yoneyama and Sebastian Scherer},
  year={2024},
  eprint={2403.03372},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/pdf/2403.03372.pdf}
}