benchmark_DEMAND_noise

This dataset is a segmented subset derived from DEMAND: Diverse Environments Multichannel Acoustic Noise Database.

It is prepared for the SPARCO noise ablation benchmark. The intended use is to provide fixed 4-second environmental noise segments for:

• AUROC-based SAE noise-related feature selection
• binary noise-presence scorer training
• scorer threshold calibration
• final held-out benchmark evaluation

Source

Original source:

• DEMAND: Diverse Environments Multichannel Acoustic Noise Database
• Authors: Joachim Thiemann, Nobutaka Ito, Emmanuel Vincent
• DOI: 10.5281/zenodo.1227121
• Source page: https://zenodo.org/records/1227121

Important license notice

The original DEMAND Zenodo page and related listings should be consulted for the authoritative license terms.

At the time this dataset card was prepared, the DEMAND license information may appear as Creative Commons Attribution 4.0 in Zenodo metadata, while other descriptions/listings may refer to Creative Commons Attribution-ShareAlike 3.0.

Because of this ambiguity, this Hugging Face dataset card uses:

license: other

Users should verify the applicable DEMAND license before redistribution, commercial use, or publishing derived mixtures. If the ShareAlike interpretation applies, derived audio segments or mixtures may need to follow compatible ShareAlike terms.

Modifications from original DEMAND

The original DEMAND recordings were processed as follows:

• only 16 kHz folders were used
• only channel 1 was used
• audio was segmented into 4-second non-overlapping clips
• incomplete trailing segments were discarded
• segments were saved as 16-bit PCM WAV files
• metadata was generated for leakage-aware split tracking

No denoising, enhancement, normalization, or artificial mixing was applied in this dataset.

Split design

The dataset uses environment-level splitting to reduce leakage risk. The same DEMAND environment never appears in more than one split.

Split	Purpose	DEMAND environments
feature_scorer_train	AUROC feature selection + noise presence scorer training	DKITCHEN, DLIVING, NFIELD, OMEETING, DWASHING, TMETRO, PRESTO, PSTATION, SPSQUARE, TCAR
scorer_val	scorer threshold calibration / checkpoint selection	OOFFICE, NPARK, TBUS
benchmark_test	final held-out evaluation only	NRIVER, OHALLWAY, STRAFFIC, PCAFETER

Recommended protocol

Use feature_scorer_train for:

• selecting noise-related SAE features using AUROC
• training the binary noise-presence scorer

Use scorer_val for:

• choosing scorer thresholds
• selecting checkpoints
• selecting intervention strength or number of edited features

Use benchmark_test only once for final evaluation.

Do not use benchmark_test for:

• AUROC feature selection
• scorer training
• threshold tuning
• selecting K
• choosing intervention strength

Files

Each row in metadata.csv corresponds to one audio segment.

Important columns:

• file_name: relative path to the audio segment
• split: one of feature_scorer_train, scorer_val, benchmark_test
• env: DEMAND environment
• channel: channel used, currently ch01
• source_group_id: leakage group, equal to environment
• temporal_group_id: environment and time range identifier
• start_sec, end_sec: segment boundaries in the original channel file
• source_file: original local source file name
• source_url: DEMAND source URL
• license_note: license caution note

Example loading

from datasets import load_dataset

ds = load_dataset("SPARCO-project/benchmark_DEMAND_noise")
print(ds)

Or locally:

from datasets import load_dataset

ds = load_dataset("audiofolder", data_dir="./benchmark_DEMAND_noise")
print(ds)

Citation

If you use this dataset, cite the original DEMAND database:

Thiemann, J., Ito, N., & Vincent, E. (2013). DEMAND: a collection of multi-channel recordings of acoustic noise in diverse environments. Zenodo. https://doi.org/10.5281/zenodo.1227121