ARCTIC-HS

An extension of the CMU_ARCTIC and L2-ARCTIC datasets for synthetic speech detection using text-to-speech, featured in the paper Synthetic speech detection with Wav2Vec 2.0 in various language settings. Specifically, the symmetric variants were used.

This dataset is 1 of 3 used in the paper, the others being:

• FLEURS-HS
  • the default train, dev and test sets
• FLEURS-HS VITS
  • test set containing (generally) more difficult synthetic samples
  • separated due to different licensing

Dataset Details

Dataset Description

The dataset features 3 parts obtained from the 2 original datasets:

• CMU (native) non-US English speakers
• CMU (native) US English speakers
• L2 (non-native) English speakers

The original ARCTIC samples are used as human samples, while synthetic samples are generated using Google Cloud Text-To-Speech.

The resulting symmetric datasets features exactly twice the samples of the original ones, but we also provide:

• human samples that couldn't be paired

  • 4 speakers in entirety we couldn't pair with a TTS voice
  • a small amount of utterances unrelated to the A and B ARCTIC samples

• synthetic samples that couldn't be paired

  • mostly when a human speaker didn't read the B ARCTIC samples

• Curated by: KONTXT by RealNetworks

• Funded by: RealNetworks

• Language(s) (NLP): English

• License: Apache 2.0 for the code, CC BY 4.0 for the dataset, however:

  • the human part of the dataset is under a custom CMU license
    • it should be compatible with CC BY 4.0
  • the human part of the L2 dataset is under CC BY-NC 4.0

Dataset Sources

The original ARCTIC sets were downloaded from their original sources.

• CMU_ARCTIC Repository: festvox.org

• L2-ARCTIC Repository: tamu.edu

• CMU_ARCTIC Paper: cmu.edu

• L2-ARCTIC Paper: tamu.edu

• Paper: Synthetic speech detection with Wav2Vec 2.0 in various language settings

Uses

This dataset is best used as a test set for accents. Each sample contains an Audio feature, and a label: human or synthetic.

Direct Use

The following snippet of code demonstrates loading the CMU non-US English speaker part of the dataset:

from datasets import load_dataset

arctic_hs = load_dataset(
    "realnetworks-kontxt/arctic-hs",
    "cmu_non-us",
    split="test",
    trust_remote_code=True,
)

To load a different part, change cmu_non-us into one of the following:

• cmu_us for CMU (native) US English speakers
• l2 for L2 (non-native) English speakers

This dataset only has a test split.

To load only the paired samples, append _symmetric to the name. For example, cmu_non-us will load the test set also containing human and synthetic samples without their counterpart, while cmu_non-us_symmetric will only load samples where there is both a human and synthetic variant. This is useful if you want to have perfectly balanced labels within speakers, and if you wish to exclude speakers for which there are no TTS counterparts at all. This is also the family of datasets used in the paper.

The trust_remote_code=True parameter is necessary because this dataset uses a custom loader. To check out which code is being ran, check out the loading script.

Dataset Structure

The dataset data is contained in the data directory.

There exists 1 directory per part.

Within those directories, there are 2 further directories:

• splits
• pairs

Within the splits folder, there is 1 file per split:

• test.tar.gz

Those .tar.gz files contain 2 directories:

• human
• synthetic

Each of these directories contain .wav files. Keep in mind that these directories can't be merged as they share most of their file names. An identical file name implies a speaker-voice pair, ex. human/arctic_a0001.wav and synthetic/arctic_a0001.wav.

The pairs folder contains a list of file names within each speaker, and whether or not there is a human-synthetic pair. Based on that metadata we determine which samples appear in symmetric datasets.

Back to the part directories, each contain 2 metadata files, which are not used in the loaded dataset, but might be useful to researchers:

• speaker-metadata.csv
  • contains the speaker IDs paired with their speech properties
• voice-metadata.csv
  • contains speaker-TTS name pairs

Finally, the data root contains a single metadata file, prompts.csv, which as the name would suggest, contains the prompt transcripts. The only samples for which there are no transcripts are the ARCTIC-C ones, for which we couldn't find a source in the internet.

Sample

A sample contains contains an Audio feature audio, and a string label.

{
  'audio': {
    'path': 'ahw/human/arctic_a0001.wav',
    'array': array([0., 0., 0., ..., 0., 0., 0.]),
    'sampling_rate': 16000
  },
  'label': 'human'
}

Citation

The dataset is featured alongside our paper, Synthetic speech detection with Wav2Vec 2.0 in various language settings, which will be published on IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). We'll provide links once it's available online.

BibTeX:

Note, the following BibTeX is incomplete - we'll update it once the actual one is known.

@inproceedings{dropuljic-ssdww2v2ivls
  author={Dropuljić, Branimir and Šuflaj, Miljenko and Jertec, Andrej and Obadić, Leo}
  booktitle={2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW)}
  title={Synthetic speech detection with Wav2Vec 2.0 in various language settings}
  year={2024}
  volume={}
  number={}
  pages={1-5}
  keywords={Synthetic speech detection;text-to-speech;wav2vec 2.0;spoofing attack;multilingualism}
  doi={}
}

Dataset Card Authors

• Miljenko Šuflaj

Dataset Card Contact

• Miljenko Šuflaj