changelinglab/cmul2arctic-l1cls

CMU + L2-ARCTIC Combined Dataset for L1 Classification

This dataset combines CMU Arctic (native English speakers) and L2-ARCTIC (non-native English speakers) for L1 (native language) classification research. It is redistributed for the purpose of academic reproducibility.

Audio is embedded in Parquet files for efficient loading via HuggingFace datasets.

Dataset Description

Subset	Source	Speakers	L1 Labels	License
cmu/	CMU Arctic	4 native English speakers	en	CMU Arctic license (see CMU Arctic website)
l2arctic/	L2-ARCTIC	24 non-native English speakers	ar, es, hi, ko, vi, zh	CC BY-NC 4.0

Splits

Split	CMU	L2-ARCTIC	Total
train	2,264	13,450	15,714
val	1,132	6,787	7,919
test	1,132	6,630	7,762
Total	4,528	26,867	31,395

Schema

Column	Type	Description
audio	struct	{path: string, bytes: binary} - Audio data with original relative path
corpus	string	cmu or l2arctic
l1_label	string	L1 language code (en, ko, zh, ar, hi, es, vi)
speaker_id	string	Speaker identifier
utt_id	string	Utterance identifier

License

⚠️ Important: This dataset contains data from two sources with different licenses.

Folder	Source	License	Commercial Use
cmu/	CMU Arctic	CMU Arctic license (see CMU Arctic website)	✅ Allowed
l2arctic/	L2-ARCTIC	CC BY-NC 4.0	❌ Non-commercial only

Combined usage requires compliance with CC BY-NC 4.0 (non-commercial use only).

Usage

Load Combined Dataset

from datasets import load_dataset, Audio

# Load one split using the pre-defined split metadata
train_ds = load_dataset("y00njaekim/cmul2arctic-l1cls", split="train")

# Or load all splits at once
ds = load_dataset("y00njaekim/cmul2arctic-l1cls")

# Access a sample
sample = train_ds[0]
print(f"L1: {sample['l1_label']}, Speaker: {sample['speaker_id']}")

# Decode audio bytes
audio_bytes = sample["audio"]["bytes"]
audio_path = sample["audio"]["path"]

Attribution

CMU Arctic

• Website: http://www.festvox.org/cmu_arctic/index.html
• Technical Report: CMU Arctic Report (PDF)

L2-ARCTIC

As requested by the original authors, please cite the following paper if you use L2-ARCTIC:

@inproceedings{zhao2018l2arctic,
    author={Guanlong {Zhao} and Sinem {Sonsaat} and Alif {Silpachai} and
            Ivana {Lucic} and Evgeny {Chukharev-Hudilainen} and John {Levis} and
            Ricardo {Gutierrez-Osuna}},
    title={L2-ARCTIC: A Non-native English Speech Corpus},
    year=2018,
    booktitle={Proc. Interspeech},
    pages={2783–2787},
    doi={10.21437/Interspeech.2018-1110},
    url={http://dx.doi.org/10.21437/Interspeech.2018-1110}
}

Original Source: https://psi.engr.tamu.edu/l2-arctic-corpus/

You can use this dataset with our benchmarking toolkit at https://github.com/changelinglab/prism

@misc{prism2026,
      title={PRiSM: Benchmarking Phone Realization in Speech Models}, 
      author={Shikhar Bharadwaj and Chin-Jou Li and Yoonjae Kim and Kwanghee Choi and Eunjung Yeo and Ryan Soh-Eun Shim and Hanyu Zhou and Brendon Boldt and Karen Rosero Jacome and Kalvin Chang and Darsh Agrawal and Keer Xu and Chao-Han Huck Yang and Jian Zhu and Shinji Watanabe and David R. Mortensen},
      year={2026},
      eprint={2601.14046},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.14046}, 
}

Acknowledgments

This dataset redistribution is provided for academic reproducibility purposes. All credit for the original data collection belongs to the respective research teams at Carnegie Mellon University and Texas A&M University.