ThaiVanPhat95/synthetic-atc-speech

Synthetic ATC Speech

Synthetic English air-traffic-control speech created for research on robust automatic speech recognition. The dataset contains 276,304 generated utterances from 15,660 unique ATC transcripts.

The dataset accompanies:

• —Contrastive Regularization for Accent-Robust ASR
• —Robust ATC ASR code
• —UWB SupCon Hybrid model
• —UWB+ATCOSIM SupCon Hybrid model

Dataset Structure

The dataset provides one training split packaged as uncompressed WebDataset TAR shards. Every sample has a paired <key>.wav audio file and <key>.json metadata record. The JSON record contains:

The dataset is balanced across four speech-generation systems and four voice-source domains:

Usage

from datasets import load_dataset

dataset = load_dataset(
    "thaivanphat95/synthetic-atc-speech",
    split="train",
    streaming=True,
)
sample = next(iter(dataset))

print(sample["wav"])
print(sample["json"]["text"])
print(sample["json"]["tts_system"])
print(sample["json"]["voice_source"])

The TAR shards are stored under data/. shard_manifest.csv records the sample count and byte size of every shard.

Intended Use

This dataset is released for non-commercial research use. It is intended for research on:

• —Robust ATC automatic speech recognition
• —Synthetic-data augmentation
• —Cross-domain and cross-dataset robustness
• —Analysis of speech-generation diversity

It should not be used as real ATC communication or relied upon in safety-critical operational systems. Commercial use is not authorized by this dataset release.

Source Data

ATC transcripts are derived from UWB-ATCC and ATCOSIM. Voice-source material is derived from four corpora:

• —UWB-ATCC: https://huggingface.co/datasets/Jzuluaga/uwb_atcc
• —ATCOSIM: https://huggingface.co/datasets/Jzuluaga/atcosim_corpus
• —L2-ARCTIC: https://psi.engr.tamu.edu/l2-arctic-corpus/
• —Marine Radio Chatter: https://www.kaggle.com/datasets/linogova/marine-radio-chatter-bridge-2-bridge-communication

Synthetic speech was generated using:

• —GLM-TTS: https://github.com/zai-org/GLM-TTS
• —CosyVoice: https://github.com/FunAudioLLM/CosyVoice
• —Afro-TTS: https://huggingface.co/intronhealth/afro-tts
• —XTTS: https://github.com/coqui-ai/TTS

Users are responsible for reviewing the terms of the source datasets, source-voice material, and speech-generation systems.

Limitations

• —Audio is synthetic and does not reproduce every property of real ATC speech.
• —Generation systems and voice sources may introduce systematic artifacts or biases.
• —Transcripts are specialized for English ATC communication.
• —The dataset may contain generation errors despite metadata integrity checks.

Citation

@article{thai2026contrastive,
  title={Contrastive Regularization for Accent-Robust ASR},
  author={Thai, Van-Phat and Dhruv, Aradhya and Pham, Duc-Thinh and Alam, Sameer},
  journal={arXiv preprint arXiv:2605.03297},
  year={2026},
  doi={10.48550/arXiv.2605.03297}
}

License

This dataset is published with license: other and is intended for non-commercial research use only. CC BY-NC 4.0 is not used because it permits non-commercial uses beyond research and may not satisfy all source-material terms, including applicable ShareAlike requirements.

Use and redistribution remain subject to the terms of the source transcripts, source voices, speech-generation systems, and related third-party material. Where those terms conflict with this research-use notice, the applicable third-party terms control. The Apache-2.0 license of the associated code repository does not cover this dataset.