AISHELL-1-Robust-54

A Mandarin speech dataset for robust automatic speech recognition (ASR), created by applying acoustic augmentation to AISHELL-1 training, development, and test data.

The augmentation workflow is the same as our LibriSpeech-based release and is inspired by Mega-ASR. This is an independently prepared dataset, not an official Mega-ASR release or a claim of exact pipeline reproduction. 54 refers to the configured acoustic scenario labels, not the number of hours.

Sources and augmentation

Source speech and reference transcripts come from AISHELL-1 / OpenSLR 33.

Background audio is drawn from the following source collections:

Source	Approximate upstream coverage
DNS3	150 noise classes; 181 hours
ESC-50	50 environmental sound classes; 2.78 hours
MUSAN	Noise: 6 hours; the corpus also includes 42.5 hours of music and 60.4 hours of speech
UrbanSound8K	10 urban sound classes; 8.75 hours

These figures describe the original source collections, not the amount sampled or the duration of this release. They do not imply that every source recording or every MUSAN component was used.

Seven base conditions are composed into individual and combined scenarios:

noise, far_field, obstructed, echo_reverb, recording, distortion, dropout

The types field preserves the complete scenario name, such as noise or far_field+distortion+noise+dropout. The + separator denotes a compound condition, not its exact processing order or parameter settings.

Format and splits

The HF configuration expects this layout. All scenarios within a split share sequentially numbered Parquet files:

README.md
train/0000.parquet, 0001.parquet, ...
dev/0000.parquet, 0001.parquet, ...
test/0000.parquet, 0001.parquet, ...

train/, dev/, and test/ map to train, validation, and test, respectively. Update the YAML paths if the uploaded directory names differ.

Field	Description
audio	Embedded WAV bytes, stored as an HF Audio feature
text	Reference Mandarin transcript
id	Source utterance ID; by default, the WAV filename without its extension
types	Full acoustic augmentation scenario name

The export does not additionally resample or transcode WAVs, or normalize transcript text. A source ID can repeat across scenarios; use (types, id) within a split to distinguish variants. Exports using --id-mode scenario instead store scenario/uid IDs. Per-split row counts and shard statistics are available in conversion_summary.json when included.

Usage

from datasets import Audio, load_dataset

# Replace with the actual repository ID.
ds = load_dataset(
    "maxwellziweiwei/AISHELL-1-Robust-54",
    split="validation",  # "train", "validation", or "test"
)

noise_ds = ds.filter(
    lambda scenario: scenario == "noise",
    input_columns=["types"],
)

# Access the original embedded WAV bytes without decoding.
raw_ds = ds.cast_column("audio", Audio(decode=False))
example = raw_ds[0]
print(example["id"], example["types"], example["text"])
wav_bytes = example["audio"]["bytes"]

Notes and licensing

• Keep all augmented versions of one source utterance in the same source partition. Do not train on development/test utterances or their variants and then report them as unseen evaluation data.
• Severe corruption may remove intelligible words while the source transcript remains unchanged. Simulation is not a substitute for evaluation on naturally recorded noisy speech.
• The four-column schema does not retain exact noise identities, effect strengths, seeds, or recording-level license records; retain generation manifests for reproducibility and attribution.
• OpenSLR 33 lists AISHELL-1 under Apache License 2.0. This does not automatically apply to all augmented mixtures: ESC-50 and UrbanSound8K include non-commercial terms; DNS3 includes multiple source licenses. Preserve applicable MUSAN and other recording-level credits as well.
• No blanket license or commercial-use clearance for the mixed audio is asserted here. Check the actual recordings used and their attribution, non-commercial, and any applicable share-alike requirements before redistribution.

Acknowledgements

We thank the creators of AISHELL-1, Mega-ASR, DNS Challenge, ESC-50, MUSAN, and UrbanSound8K. Please acknowledge the original corpora and the augmentation method in publications:

• Bu et al., AISHELL-1: An Open-Source Mandarin Speech Corpus and a Speech Recognition Baseline, 2017.
• Xie et al., Mega-ASR: Towards In-the-wild² Speech Recognition via Scaling up Real-world Acoustic Simulation, 2026.