VoiceAeroRef

VoiceAeroRef is an audio-guided referring image segmentation resource for aerial and remote-sensing imagery. It pairs RISBench image-expression-mask triplets with spoken English referring expressions and provides controlled speech-corruption splits for robustness evaluation.

Dataset summary

The release contains 262,696 audio-metadata records in 44 deterministic WebDataset TAR shards (approximately 45.05 GB). Audio is stored as 16 kHz, mono, PCM16 WAV.

Configuration	Split	Records	Voice policy
clean	train	210,352	eight voices per valid source
clean	validation	10,013	one balanced voice per expression
clean	test	16,159	one balanced voice per expression
hard	validation	10,013	one controlled corruption per expression
hard	test	16,159	one controlled corruption per expression

The source audit starts from 52,472 RISBench triplets. Six training sources whose expression contains only punctuation are excluded, leaving 52,466 valid image-expression-mask triplets.

Speech inventory

Clean speech uses eight English neural voices:

Locale	Accent	Female voice	Male voice
en-US	US	Jenny	Guy
en-GB	GB	Sonia	Ryan
en-AU	AU	Natasha	William
en-IN	IN	Neerja	Prabhat

For training, every valid source has all eight voices. The intended training protocol samples exactly one voice for each source in each epoch. Validation and test use one deterministically balanced voice per expression.

Hard configuration

The hard validation and test splits use three synthetic interference types:

• rotor noise;
• wind noise;
• mixed rotor and wind noise.

Each type has low, medium and high severity, centred at 15 dB, 7.5 dB and 0 dB SNR respectively, with deterministic per-utterance perturbation. Training speech remains clean.

Record format

Every WebDataset example contains a WAV file and a JSON file with the same key:

test_000000__au_female.wav
test_000000__au_female.json

The JSON metadata contains fields such as:

{
  "sample_id": "test_000000",
  "source_index": 0,
  "image": "img_rgb/test_0_0.png",
  "mask": "mask/test_0_0.png",
  "phrase": "The baseball field featured in the image ...",
  "voice": "en-AU-NatashaNeural",
  "accent": "AU",
  "gender": "female",
  "locale": "en-AU",
  "voice_id": "au_female",
  "config": "clean",
  "split": "test",
  "condition": "clean"
}

Hard examples additionally contain noise_type, severity, snr_db and noise_seed.

Loading with 🤗 Datasets

Streaming avoids downloading the complete release:

from datasets import load_dataset

dataset = load_dataset(
    "lironui/VoiceAeroRef",
    "clean",
    streaming=True,
)
sample = next(iter(dataset["train"]))
print(sample.keys())

Hard evaluation data:

hard = load_dataset(
    "lironui/VoiceAeroRef",
    "hard",
    split="test",
    streaming=True,
)
sample = next(iter(hard))

Downloading files

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download the complete repository:

hf download lironui/VoiceAeroRef \
  --repo-type dataset \
  --local-dir VoiceAeroRef

Download only clean test shards and release metadata:

hf download lironui/VoiceAeroRef \
  --repo-type dataset \
  --include "data/clean/test-*.tar" \
  --include "release_manifest.json" \
  --include "checksums.sha256" \
  --local-dir VoiceAeroRef

The companion AeroReformer2 repository includes scripts/prepare_voiceaeroref.py, which downloads selected configurations, checks optional SHA-256 digests, extracts WAV files and writes training-ready JSONL manifests.

Images and masks

This repository distributes speech and metadata only. RISBench imagery and masks are not duplicated. The image and mask fields preserve the original RISBench references; users must obtain RISBench separately and comply with its terms.

Integrity and exclusions

• release_manifest.json records every shard, record count, byte size and SHA-256 digest.
• checksums.sha256 provides a checksum list suitable for independent verification.
• excluded_samples.jsonl documents the six invalid punctuation-only source expressions.

Intended use

VoiceAeroRef supports research on:

• audio-guided referring segmentation;
• aerial and remote-sensing scene understanding;
• speech robustness under rotor and wind interference;
• accent- and gender-aware evaluation.

It is not designed for speech recognition benchmarking, speaker identification, surveillance decisions or safety-critical deployment.

Limitations

• Speech is synthesized and does not reproduce the full variability of human speech.
• The release contains English speech with four accent groups.
• Corruptions are controlled synthetic conditions rather than recordings from every real flight environment.
• Dataset performance depends on the coverage and annotation properties of RISBench.

Licensing and attribution

This dataset card does not grant additional rights to RISBench imagery or annotations. Users and redistributors must comply with the original RISBench terms and with any terms governing the speech-generation service. Before public release, the publisher should add the final dataset license selected after verifying compatibility with those upstream terms.

Citation

@dataset{voiceaeroref2026,
  title   = {VoiceAeroRef: Audio-Guided Referring Segmentation for Aerial Imagery},
  author  = {AeroReformer2 Authors},
  year    = {2026},
  url     = {https://huggingface.co/datasets/lironui/VoiceAeroRef}
}