Datasets:

changelinglab
/

speechocean-l2eval

license: cc-by-nc-4.0
arxiv: 2601.14046
dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: speaker_id
      dtype: string
    - name: utt_id
      dtype: string
    - name: text
      dtype: string
    - name: accuracy
      dtype: int32
    - name: completeness
      dtype: float32
    - name: fluency
      dtype: int32
    - name: prosodic
      dtype: int32
    - name: total
      dtype: int32
  splits:
    - name: train
      num_bytes: 260979874
      num_examples: 2260
    - name: val
      num_bytes: 37136358
      num_examples: 240
    - name: test
      num_bytes: 288161567
      num_examples: 2500
  download_size: 610453123
  dataset_size: 586277799
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
      - split: test
        path: data/test-*
task_categories:
  - automatic-speech-recognition
language:
  - en
size_categories:
  - 1K<n<10K

speechocean762: A non-native English corpus for pronunciation scoring task

Dataset Summary

speechocean762 is an open-source non-native English speech corpus designed for pronunciation assessment and L2 spoken proficiency modeling. This Hugging Face version provides sentence-level audio and expert scores, organized into standard train / validation / test splits.

All speakers are Mandarin L1 learners of English, spanning both children and adults. Each utterance is evaluated independently by five expert annotators using standardized pronunciation metrics.

This dataset is suitable for:

pronunciation scoring
L2 speech assessment
speech representation learning
downstream regression or classification tasks

Dataset Structure

Splits

The dataset is published with three predefined splits:

train (2260)
val (240)
test (2500)

Splits are speaker-disjoint and provided as native Hugging Face splits.

Features

Each example contains:

Field	Type	Description
`audio`	`Audio`	Speech waveform (16 kHz)
`speaker_id`	`string`	Speaker identifier
`utt_id`	`string`	Utterance identifier
`text`	`string`	Prompt sentence
`accuracy`	`int`	Sentence-level pronunciation accuracy
`completeness`	`float`	Percentage of correctly pronounced words
`fluency`	`int`	Sentence-level fluency score
`prosodic`	`int`	Sentence-level prosody score
`total`	`int`	Overall pronunciation score

Scoring Metrics (Sentence level)

All sentence-level scores follow the original speechocean762 definitions. For detailed descriptions, see:

arXiv: https://arxiv.org/abs/2104.01378
Github: https://github.com/jimbozhang/speechocean762

Dataset Creation

This Hugging Face dataset is derived from the original speechocean762 corpus and includes:

sentence-level audio
sentence-level expert scores
standardized HF Audio features
speaker-disjoint train/val/test splits

Word-level and phoneme-level annotations are not included in this version.

Source Dataset: https://huggingface.co/datasets/mispeech/speechocean762

License

The original speechocean762 dataset is released for free use, including commercial and non-commercial purposes, as stated by the original authors. Users should consult the original repository for full licensing details.

Citation

If you use this dataset, please cite the original paper:

@inproceedings{zhang2021speechocean762,
  title={speechocean762: An Open-Source Non-native English Speech Corpus For Pronunciation Assessment},
  author={Zhang, Junbo and Zhang, Zhiwen and Wang, Yongqing and Yan, Zhiyong and Song, Qiong and Huang, Yukai and Li, Ke and Povey, Daniel and Wang, Yujun},
  booktitle={Proc. Interspeech 2021},
  year={2021}
}

Acknowledgements

All credit for data collection and annotation belongs to the original speechocean762 authors. This Hugging Face release focuses on standardized access and reproducibility for modern speech and representation learning pipelines.

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}, 
}