benchmark-noise_ablation

This dataset contains fixed-length noisy speech mixtures generated from LibriSpeech test-clean clean speech and two interference types:

• 50% DEMAND environmental background noise
• 50% generated white noise

The dataset is intended for SPARCO noise ablation experiments, including:

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

Sources

Clean speech:

• LibriSpeech ASR corpus, test-clean split
• Source: https://www.openslr.org/12
• License: CC BY 4.0 according to OpenSLR

Background noise:

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

White noise:

• Generated programmatically using a fixed random seed.

Important license notice

This dataset is a derived dataset containing mixtures of LibriSpeech and DEMAND audio.

LibriSpeech is listed by OpenSLR under CC BY 4.0. DEMAND license information should be verified from the original DEMAND Zenodo page and associated documentation. DEMAND metadata or documentation may refer to Creative Commons Attribution 4.0 and/or Creative Commons Attribution-ShareAlike 3.0 Unported. Because of this ambiguity, this dataset card conservatively uses:

license: other

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

Generation configuration

• clean speech source: LibriSpeech test-clean only
• sample rate: 16000 Hz
• clip length: 4.0 seconds
• DEMAND/white ratio: approximately 50% / 50% per split
• SNR choices in dB: 5, 10, 15, 20
• SNR assignment: stratified balanced over noise_type × SNR within each split
• clean speech split: speaker-disjoint
• DEMAND split: inherited from benchmark_DEMAND_noise

For the recommended SNR set [5, 10, 15, 20], each split is balanced over eight cells:

Noise type	5 dB	10 dB	15 dB	20 dB
DEMAND	balanced	balanced	balanced	balanced
White noise	balanced	balanced	balanced	balanced

Splits

Split	Purpose
feature_scorer_train	AUROC feature selection + noise-presence scorer training
scorer_val	scorer threshold calibration / checkpoint selection
benchmark_test	final held-out evaluation only

Do not use benchmark_test for feature selection, scorer training, threshold tuning, selecting K, or selecting intervention strength.

Files

• metadata.csv: one row per mixture
• scorer_examples.csv: binary noise-presence scorer examples
• mixtures/: noisy speech mixtures
• clean/: clean speech targets cropped/padded using the same 4-second window as the mixture

Important metadata columns:

• file_name: relative path to mixture audio
• clean_file_name: relative path to clean target audio
• split: feature_scorer_train, scorer_val, or benchmark_test
• noise_type: demand or white
• snr_db: target SNR in dB
• speaker_id: LibriSpeech speaker id
• noise_env: DEMAND environment for demand noise, white for white noise
• noise_segment_id: DEMAND segment id for demand noise, generated id for white noise

Loading locally

from datasets import load_dataset

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

Loading from Hugging Face

from datasets import load_dataset

ds = load_dataset("SPARCO-project/benchmark-noise_ablation")
print(ds)

Citations

Please cite LibriSpeech and DEMAND if you use this dataset.

• LibriSpeech: Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, 2015.
• DEMAND: 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