Hugging Face's logo

Datasets:
yairamr
/
voicebank-demand-fingerprint-16k

Tasks:
Audio-to-Audio
Audio Classification
Modalities:
Audio
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
Tags:
Synthetic
DOI:
Libraries:
Datasets
Dask
Dataset card Data Studio Files
xet
Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Enhanced VoiceBank+DEMAND Dataset with Noise Fingerprints (16000Hz)

Dataset Description

⚠️ Important Note: This is NOT the exact same dataset as the original VoiceBank-DEMAND dataset. While we follow the VoiceBank-DEMAND specification for speaker splits and SNR levels, the noise types used are different due to DEMAND dataset availability and environmental focus.

This dataset contains enhanced noisy speech samples created by mixing clean speech from the VCTK corpus with environmental noise from the DEMAND dataset, with added noise fingerprints for advanced research.

Dataset Summary

  • Source: VCTK corpus (clean speech) + DEMAND dataset (noise)
  • Sample Rate: 16000 Hz
  • SNR Levels: Uniform distribution over [-5, 20] dB
  • Noise Types: 15 environmental noise types (10 train + 5 test)
  • Test Speakers: p226, p287 (VoiceBank-DEMAND specification)
  • Train Speakers: 28 speakers (VoiceBank-DEMAND specification)
  • Total Samples: 11,824 (11,044 train + 780 test)
  • Audio Files: Included in dataset

🚀 Enhanced Features

🎯 Noise Fingerprints

  • 1-second noise segments extracted separately from mixing
  • Enables advanced noise identification and analysis

🏷️ Original DEMAND Noise Names

  • Preserves original DEMAND noise naming (e.g., 'DKITCHEN', 'TCAR', 'SPSQUARE')
  • No mapping applied - maintains dataset consistency

📝 Enhanced Metadata Structure

  • Nested directory structure with comprehensive metadata
  • Standardized ASR labels and noise type preservation
  • Individual text files per sample for easy access

Dataset Structure

Each sample contains:

  • noisy: Noisy speech audio (clean + noise mixed at specified SNR)
  • clean: Clean speech audio normalized to -25 dBFS
  • noise_fp: 🎯 1-second noise fingerprint audio (separate from noisy signal)
  • text: Text transcription of the speech
  • noise_type: Original DEMAND noise name (e.g., "DKITCHEN", "TCAR")
  • snr: Signal-to-noise ratio used for mixing (in dB)
  • id: Unique sample identifier
  • speaker: Speaker ID from VCTK corpus

Noise Type Configuration

Train Noise Types (10):

  • DKITCHEN, OOFFICE, TCAR, TMETRO, STRAFFIC, TBUS, NPARK, NFIELD, DLIVING, DWASHING

Test Noise Types (5):

  • SPSQUARE, PRESTO, PSTATION, NRIVER, PCAFETER

Audio Processing Pipeline

  1. Clean Speech: VCTK corpus resampled to 16000Hz and normalized to -25 dBFS
  2. Noise Selection: DEMAND dataset files filtered by train/test split and resampled to 16000Hz
  3. Mixing: Clean speech mixed with noise at VoiceBank-DEMAND SNR levels
  4. 🎯 Noise Fingerprint: 1-second segment extracted separately from noise file (not used in mixing)
  5. Metadata: Enhanced structure with original DEMAND noise names and standardized labels

Usage 

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("yairamr/voicebank-demand-fingerprint-16k")

# Access a sample
sample = dataset["train"][0]
print(f"Noisy audio shape: {sample['noisy']['array'].shape}")
print(f"Clean audio shape: {sample['clean']['array'].shape}")
print(f"Noise fingerprint shape: {sample['noise_fp']['array'].shape}")
print(f"Transcription: {sample['text']}")
print(f"Noise type: {sample['noise_type']}")
print(f"SNR: {sample['snr']} dB")
print(f"Speaker: {sample['speaker']}")

# Play audio
import soundfile as sf
sf.write('noisy.wav', sample['noisy']['array'], 16000)
sf.write('clean.wav', sample['clean']['array'], 16000)
sf.write('noise_fp.wav', sample['noise_fp']['array'], 16000)

🔍 Key Differences from Original VoiceBank-DEMAND

Aspect Original VoiceBank-DEMAND This Enhanced Version
Speaker Split ✅ p226, p287 (test) ✅ p226, p287 (test)
SNR Levels ✅ Train [15,10,5,0], Test [17.5,12.5,7.5,2.5] ⚠️ Modified Uniform [-5, 20] dB
Noise Types ⚠️ Different selection ⚠️ Modified selection (15 environmental types)
Structure Flat directories Enhanced: Nested directories
Audio Files noisy.wav, clean.wav ⚠️ Enhanced: + noise_fp.wav (fingerprints)
Metadata Basic files ⚠️Enhanced: Individual txt files per sample

Citation 

@dataset{voicebank_demand_fingerprint_16k,
  title={voicebank-demand with noise fingerprint},
  author={Yair Amar},
  year={2025},
  url={https://huggingface.co/datasets/yairamr/voicebank-demand-fingerprint-16k}
}

License

This dataset is released under the MIT License. The underlying VCTK corpus and DEMAND dataset have their own licenses.

Acknowledgments

  • VCTK Corpus: Yamagishi, Junichi; Veaux, Christophe; MacDonald, Kirsten (2019)
  • DEMAND Dataset: Thiemann, Joachim; Ito, Nobutaka, Vincent, Emmanuel (2013)
  • This Enhanced Version: Enhanced VoiceBank+DEMAND with Noise Fingerprints

⚠️ Important Dataset Notes

  1. This is NOT the exact original VoiceBank-DEMAND dataset - noise types are different
  2. Speaker splits match VoiceBank-DEMAND specification - p226, p287 for test
  3. SNR levels use uniform distribution - [-5, 20] dB for both train and test
  4. Enhanced with noise fingerprints - 1-second segments for advanced research
  5. Environmental noise focus - no speech-like or artificial sounds

🛠️ Created with

  • Random seed: 42
  • Final sample count: 11,824 (11,044 train + 780 test)
  • VoiceBank-DEMAND compliance: Speaker splits ✅
  • Original DEMAND names: Preserved ✅
Downloads last month
6

Collection including yairamr/voicebank-demand-fingerprint-16k