CASPER: A Large Scale Spontaneous Speech Dataset

CASPER is a speech dataset comprised of recordings of two people engaging in spontaneous conversations in English. The dataset aims to fill the gap in high quality spontaneous speech data. The conversations were conducted over a custom-built web platform from each participant's end and their own device, fostering genuine interactions and natural conversations with a diverse range of topics.

How to Download the Dataset

Request Access

At this point in time, the dataset is open for research usage. Due this and reasons related to licensing you need to ask for access using your HF account before you are able to download the dataset.

Option 1: Clone the Repository

You can clone the repository into your local directory with all the files in the dataset.

# Install git lfs
git lfs install

# Clone the dataset, you should have your HF access token initialized
git clone https://huggingface.co/datasets/CASPER-SSSD/CASPER

Option 2: Clone the Repository

You can download the snapshot of the dataset, which does not create a git repository but simply downloads everything in the dataset.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="CASPER-SSSD/CASPER",
    repo_type="dataset",
    local_dir="path/to/local/directory",
    local_dir_use_symlinks=False
)

Dataset Format

The dataset is split into 2 splits, train and test.

Recordings

Each split contains folders with 2 .wav audio files in them. The name of each folder is the recording ID, and each .wav file is from one participant's end. For example, the following recording has ID "2433_173715_UtOU_UF3d_t9s", the first file contains speech from the speaker "F3dDFVMq7uNk9eoyr4gToZ8viX23", and the second one from the other speaker "tOUe12sveSe2zOTxV7xBSqDa5Eg2":

train/
├── recordings/
│   ├── 2433_173715_UtOU_UF3d_t9s/
│   │   ├── 2433_173715_UtOU_UF3d_t9s-F3dDFVMq7uNk9eoyr4gToZ8viX23-0.wav
│   │   └── 2433_173715_UtOU_UF3d_t9s-tOUe12sveSe2zOTxV7xBSqDa5Eg2-0.wav
│   └── ...

Metadata

The metadata about the recordings (the speakers in a recording, topic etc.) and the speakers (age, accent, etc.) are shared in the metadata.tar.gz file. After unzipping, you will find two folders named csv and json under both the train and test folders. Each folder contains metadata in the respective format, both of them carry the same information.

There are 3 files about metadata: recordings.csv/json, speakers.csv/json, topics.csv/json

recordings.csv/json has recording ID (e.g. 2433_173715_UtOU_UF3d_t9s) as the primary index, and links to the access codes of speakers in that recording, as well as the topic ID of the recording.

speakers.csv/json has metadata about speakers with speaker access code (e.g. F3dDFVMq7uNk9eoyr4gToZ8viX23) as the primary index, and links to the metadata of the speaker (e.g. age, accent).

topics.csv/json has the topic ID (e.g. 101) as the primary index and links to the name and description of the topic (e.g. fav-activity, What is your favorite activity to do in your free time?)

Contact Us

You can reach out to us through this e-mail address: [email protected]

License Disclaimer

This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. By accessing or using this dataset, you agree not to use the data, directly or indirectly, to identify or re-identify individuals contained within the dataset. Any attempt to identify individuals or violate their privacy is strictly prohibited.