dNoize

Meaningful Speech Enhancement (Noise and Reverb Suppression) | Native-Friendly to torch_directml on Windows 10/11

VoiceBank-DEMAND Dataset Setup

This guide explains how to download and prepare the VoiceBank-DEMAND dataset for audio denoising training.

Dataset Information

Source: University of Edinburgh DataShare
URL: https://datashare.ed.ac.uk/handle/10283/2791

The VoiceBank-DEMAND dataset contains clean and noisy speech recordings for training speech enhancement models.

Downloads

Download the following files from the dataset page:

Training Data (28 speakers)

• clean_trainset_28spk_wav.zip (2.315 GB)
• noisy_trainset_28spk_wav.zip (2.635 GB)

Test Data

• clean_testset_wav.zip (147.1 MB)
• noisy_testset_wav.zip (162.6 MB)

Optional: Larger Training Set (56 speakers)

• clean_trainset_56spk_wav.zip (4.442 GB)
• noisy_trainset_56spk_wav.zip (5.240 GB)

Directory Structure

After downloading, organize the files in the following structure:

project_root/
├── data/
│   ├── train/
│   │   ├── clean_trainset_28spk_wav/
│   │   │   ├── p226_001.wav
│   │   │   ├── p226_002.wav
│   │   │   └── ...
│   │   └── noisy_trainset_28spk_wav/
│   │       ├── p226_001.wav
│   │       ├── p226_002.wav
│   │       └── ...
│   └── test/
│       ├── clean_testset_wav/
│       │   ├── p232_001.wav
│       │   ├── p232_002.wav
│       │   └── ...
│       └── noisy_testset_wav/
│           ├── p232_001.wav
│           ├── p232_002.wav
│           └── ...
└── ...

Setup Instructions

Step 1: Create Directory Structure

mkdir -p data/train
mkdir -p data/test

Step 2: Extract Training Data

# Extract clean training audio
unzip clean_trainset_28spk_wav.zip -d data/train/

# Extract noisy training audio
unzip noisy_trainset_28spk_wav.zip -d data/train/

Step 3: Extract Test Data

# Extract clean test audio
unzip clean_testset_wav.zip -d data/test/

# Extract noisy test audio
unzip noisy_testset_wav.zip -d data/test/

Dataset Statistics

Training Set (28 speakers)

• Clean files: 11,572 audio files
• Noisy files: 11,572 audio files
• Total speakers: 28
• Sample rate: 48 kHz (will be resampled to 16 kHz during training)

Test Set

• Clean files: 824 audio files
• Noisy files: 824 audio files
• Total speakers: 2
• Sample rate: 48 kHz (will be resampled to 16 kHz during inference)

File Naming Convention

Files are paired by name:

• Clean: data/train/clean_trainset_28spk_wav/p226_001.wav
• Noisy: data/train/noisy_trainset_28spk_wav/p226_001.wav

The naming format is p{speaker_id}_{utterance_id}.wav

Troubleshooting

Issue: Sample rate errors

The dataset uses 48 kHz audio, but the model expects 16 kHz. This is handled automatically during training:

# In load_audio function (already implemented):
if sr != 16000:
    resampler = torchaudio.transforms.Resample(sr, 16000)
    wave = resampler(wave)

Issue: Extraction errors on Windows

Use 7-Zip or WinRAR if the built-in Windows extractor fails:

Right-click zip file -> 7-Zip -> Extract to "data/train/"

License and Citation

If you use this dataset, please cite:

Valentini-Botinhao, C. (2017). Noisy speech database for training speech enhancement 
algorithms and TTS models [Data set]. University of Edinburgh. School of Informatics. 
Centre for Speech Technology Research (CSTR). https://doi.org/10.7488/ds/2117

Additional Requirements

Before installing the python packages using requirements.txt, you need to download and install Windows 11 SDK (latest version) using the Visual Studio Installer Tool. This is required for building pesq package. If you are not going to use pesq, you can skip it but do not forget to remove it from the train code.