The WSJ0 Hipster Ambient Mixtures (WHAM!) dataset pairs each two-speaker mixture in the wsj0-2mix dataset with a unique noise background scene. We also created WHAMR!, an extension that adds artificial reverberation to the speech signals in addition to the background noise.

The noise audio was collected at various urban locations throughout the San Francisco Bay Area in late 2018. The environments primarily consist of restaurants, cafes, bars, and parks. Audio was recorded using an Apogee Sennheiser binaural microphone on a tripod between 1.0 and 1.5 meters off the ground.

The set of noise samples, referred to as "WHAM! noise dataset", is provided here, along with the scripts to build the WHAM! and WHAMR! datasets from the noise data and the WSJ0 dataset. We also provide the "WHAM!48kHz noise dataset", consisting of the noise recordings at their original sample rate and without segmenting the clips to the duration of WSJ0 clips. Both the WHAM! noise dataset and the WHAM!48kHz noise dataset have been processed to remove any segments containing intelligible speech. Because the WHAM!48kHz noise dataset has not been further segmented to the duration of WSJ0 clips, it has a wide distribution of clip durations.

This work is further described in our papers "WHAM!: Extending Speech Separation to Noisy Environments." and "WHAMR!: Noisy and Reverberant Single-Channel Speech Separation."

Download

The WHAM! noise dataset and variants, along with relevant data generation scripts are available for download:

• WHAM! noise dataset. (17.65 GB unzipping to 35 GB) MD5: 8649eafcd6dc391fda8deab7cc6b662e
• Original WHAM! data generation scripts and documentation. (674 KB unzipping to 2.8 MB) MD5: c836e2fce9449ee0d063c15c2fce0fb0
• WHAMR! data generation scripts and documentation. (4.8 MB unzipping to 11.6 MB) MD5: 11a2384408bab4b7f3c64f171a593c70
• WHAM!48kHz noise dataset: Full-length noise recordings at the original sample rate. (68.1 GB unzipping to 76 GB) MD5: b11cff68963f24acdefc64aa42766fa2

Building the Dataset

The WHAM! dataset is built by mixing 2-speaker mixtures from the wsj0-2mix dataset with noise samples from the WHAM! noise dataset. Only the noise data is provided here, and users will need access (and license) to the WSJ0 dataset.

For WHAM!: Please refer to the README for detailed instructions on how to use the mixing scripts, which can be downloaded using the link above.

For WHAMR!: Please refer to the README for detailed instructions on how to use the mixing scripts, which can be downloaded using the link above.

Dataset Structure

The WHAM! noise dataset is split into training, validation, and test sets following the wsj0-2mix dataset.

The clips are in 32-bit floating point WAV format with 2 channels and a sampling rate of 16 kHz. The average clip duration is 10 seconds with the shortest clip being 3.5 seconds and the longest 47.7 seconds.

Citation

WHAM! is a joint effort between Mitsubishi Electronics Research Laboratories (MERL) and Whisper. If you use WHAM! or WHAM!48kHz please cite our paper describing the dataset:

@inproceedings{Wichern2019WHAM,
    title     = {WHAM!: Extending Speech Separation to Noisy Environments},
    author    = {Wichern, Gordon and Antognini, Joe and Flynn, Michael and Zhu,
                 Licheng Richard and McQuinn, Emmett and Crow,
                 Dwight and Manilow, Ethan and Le Roux, Jonathan},
    booktitle = {Proc. Interspeech},
    year      = {2019},
    month     = sep
}

If you use WHAMR! please cite our paper describing the dataset:

@inproceedings{Maciejewski2020WHAMR,
    title     = {WHAMR!: Noisy and Reverberant Single-Channel Speech Separation},
    author    = {Maciejewski, Matthew and Wichern, Gordon and Le Roux, Jonathan},
    booktitle = {Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
    year      = {2020},
    month     = may
}