Voicebank-Demand Dataset

Voicebank-Demand, also known as VCTK-Demand, is a noise suppression dataset with a sampling rate of 48kHz.
There are two train datasets: one is a 28-speaker version, and the other is a 56-speaker version.
In many papers, including ours, the 28-speaker version is used.

Preparing dataset

Download

Download the train data, test data, and logfiles from here.
Download a trainscript file of the testset from here.

Downsample

If needed, downsample the dataset using scripts/resample.py.
For example, if you want to downsample to 16kHz, run the code below:

python -m scripts.resample --to-sr 16000 --from-dir ~/Datasets/voicebank-demand/48k --to-dir ~/Datasets/voicebank-demand/16k

After downloading, the directory may look like this:

voicebank-demand
├─ 16k
|  ├─ clean_testset_wav
|  ├─ clean_trainset_28spk_wav
|  ├─ noisy_testset_wav
|  └─ noisy_trainset_28spk_wav
├─ 48k
|  ├─ clean_testset_wav
|  ├─ clean_trainset_28spk_wav
|  ├─ noisy_testset_wav
|  └─ noisy_trainset_28spk_wav
└─ logfiles
   ├─ log_readme.txt
   ├─ log_testset.txt
   ├─ log_trainset_28spk.txt
   └─ transcript_testset.txt

Modify Configuration file

You have to change the dataset path and sampling rate of configuration file. For example, to train FastEnhancer-B, change data section in configs/fastenhancer/b.yaml.

Dataset Code

For Voicebank-demand dataset, we load clean and noisy speech pairs.
To check or modify the dataset code, see utils/data/voicebank_demand.py.

Training Code

To check or modify the training code for Voicebank-Demand, see wrappers/ns.py.
Clean and noisy pairs are loaded for training.