nkdem/DNS-Challenge-2020-DevTest-16k

DNS Challenge 2020 Dev Test Set

Preprocessed dev test set from the Interspeech 2020 DNS Challenge.

Dataset Structure

Split	Samples	Clean Reference	Description
synthetic_no_reverb	150	✓	Anechoic synthetic mixtures
synthetic_with_reverb	150	✓	Reverberant synthetic mixtures
real_recordings	300	✗	Real-world noisy recordings

Usage

from datasets import load_dataset
ds = load_dataset("nkdem/DNS-Challenge-2020-DevTest-16k")

# Access a sample
sample = ds['synthetic_no_reverb'][0]
print(sample.keys())  # ['id', 'fileid', 'noisy', 'clean', 'subset', 'has_reference']

Audio Format

• 16kHz mono WAV
• 10 seconds per clip
• Stored as bytes (use soundfile to decode)

import soundfile as sf
import io

audio, sr = sf.read(io.BytesIO(sample['noisy']['bytes']), dtype='float32')

Source

This dataset is derived from the official DNS Challenge 2020 test set. The original data was downloaded from Microsoft's DNS Challenge repository and repackaged for easier use with HuggingFace datasets.

Citation

@inproceedings{reddy2020interspeech,
  title={The INTERSPEECH 2020 Deep Noise Suppression Challenge: Datasets, Subjective Testing Framework, and Challenge Results},
  author={Reddy, Chandan KA and Gopal, Vishak and Cutler, Ross and Beeseley, Ebrahim and Aichner, Robert and Braun, Sebastian and Gamper, Hannes and Haas, Sebastian and Kindermans, Pieter-Jan and Loide, Marko and others},
  booktitle={Interspeech},
  year={2020}
}