LibriSpeech-Robust-54

Multi-Condition Acoustic Augmentation for Robust ASR

An English speech dataset for training and evaluating automatic speech recognition (ASR) under simulated acoustic degradations, covering training, development, and test data. Source speech and reference transcripts originate from LibriSpeech. Augmented recordings cover individual and combined conditions such as background noise, far-field effects, obstruction, reverberation, recording artifacts, distortion, and dropout.

The 54 in the name refers to the configured acoustic scenario labels, not the number of hours, languages, or source subsets. Exact exported coverage and row counts should be checked separately for each split.

The augmentation design is inspired by Mega-ASR: Towards In-the-wild² Speech Recognition via Scaling up Real-world Acoustic Simulation. This is an independently prepared dataset, not the official Mega-ASR training dataset or benchmark. It does not claim an exact reproduction of the original synthesis pipeline or parameter distributions.

This repository contains augmented speech with ASR transcripts, rather than a standalone collection of noise recordings.

Data sources

Source speech and transcripts

LibriSpeech provides the English read-speech recordings and reference transcripts for all three partitions. Augmented training, development, and test recordings are exposed as the Hugging Face train-clean-100, train-clean-360, train-other-500, dev-clean, dev-other,and test-clean , test-other splits, respectively.

Noise and background-audio sources

Source collection	Upstream coverage and approximate duration	Reference
DNS Challenge 3 (DNS3)	Approximately 150 noise classes; 181 hours of noise	Official DNS3 repository, DNS3 paper, §3.2
ESC-50	50 environmental sound classes; 2,000 five-second clips, approximately 2.78 hours	Official repository
MUSAN	Approximately 6 hours of noise; the source corpus also contains approximately 42.5 hours of music and 60.4 hours of speech	OpenSLR 17, MUSAN paper
UrbanSound8K	10 urban sound classes; 8,732 excerpts, approximately 8.75 hours	Official dataset page, original paper

Acoustic conditions

The scenario vocabulary uses seven base labels:

Label	Acoustic condition
noise	Additive background noise or background-audio interference
far_field	Simulated distant-microphone acoustic degradation
obstructed	Simulated obstruction or occlusion of the acoustic path
echo_reverb	Echo and reverberation effects
recording	Recording or channel artifacts
distortion	Signal distortion effects
dropout	Temporal audio loss or interruption

Splits and repository structure

All scenarios within each split are stored together in sequential Parquet shards under a single default configuration. The README configuration expects the following repository layout:

.
├── README.md
├── train-clean-100/
│   ├── 0000.parquet
│   ├── 0001.parquet
│   └── ...
├── dev-clean/
│   ├── 0000.parquet
│   ├── 0001.parquet
│   └── ...
└── test-clean/
    ├── 0000.parquet
    ├── 0001.parquet
    └── ...

Usage

Load the dataset

from datasets import load_dataset

repo_id = "maxwellziweiwei/LibriSpeech-Robust-54"

# Use "train", "validation", or "test".
ds = load_dataset(
    repo_id,
    split="validation",
)

print(ds.features)
print(len(ds))

For a large training split, streaming avoids downloading the entire split in advance:

train_stream = load_dataset(repo_id, split="train", streaming=True)

Select one acoustic condition

# Read only the types column when filtering; audio decoding is unnecessary.
noise_ds = ds.filter(
    lambda scenario: scenario == "noise",
    input_columns=["types"],
)

Exact equality selects one scenario. To select every scenario containing the noise component:

all_noise_ds = ds.filter(
    lambda scenario: "noise" in scenario.split("+"),
    input_columns=["types"],
)

Access the embedded WAV bytes

from datasets import Audio

raw_ds = ds.cast_column("audio", Audio(decode=False))
example = raw_ds[0]

print(example["id"])
print(example["types"])
print(example["text"])
print(example["audio"]["path"])
wav_bytes = example["audio"]["bytes"]

Acknowledgements and references

We thank the authors and maintainers of Mega-ASR, LibriSpeech, DNS Challenge, ESC-50, MUSAN, and UrbanSound8K, together with the original recording contributors.

Please cite the source corpora and the method that inspired the acoustic simulation:

• Mega-ASR: Towards In-the-wild² Speech Recognition via Scaling up Real-world Acoustic Simulation, Xie et al., 2026.
• LibriSpeech: An ASR corpus based on public domain audio books, Panayotov et al., ICASSP 2015.
• INTERSPEECH 2021 Deep Noise Suppression Challenge, Reddy et al., 2021.
• ESC: Dataset for Environmental Sound Classification, Piczak, ACM Multimedia 2015.
• MUSAN: A Music, Speech, and Noise Corpus, Snyder, Chen, and Povey, 2015.
• A Dataset and Taxonomy for Urban Sound Research, Salamon, Jacoby, and Bello, ACM Multimedia 2014.