You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

SS-JDSC: 単一話者による日本語構音障害音声コーパス / Single-speaker Japanese dysarthric speech corpus

本コーパスは,日本語の構音障害音声認識およびその関連タスクのために設計された,単一話者による読み上げ音声コーパスです.鼻咽腔閉鎖不全に伴う構音障害音声を約15時間分含んでいます.

SS-JDSC is the corpus of Japanese dysarthric speech designed for research on automatic speech recognition and related tasks. It provides 15 hours of speech data from a single Japanese male speaker with velopharyngeal insufficiency (VPI).

コーパスデザイン / Corpus design

本コーパスは 3 つのサブセットから成ります.

  • basic: 通常音声(JSUTコーパス BASIC5000)と同じ文を発話./ Parallel to control speech in JSUT BASIC5000.
  • hard: 聞き取り困難な音素を含む発話./ Phoetically confusable utterances.
  • daily: 日常会話に使われるフレーズを含む発話./ Daily-use phrases.

音声認識評価のために,それぞれのサブセットについてテストセットリスト(test_{basic, hard, daily}.txt)を含んでいます.音声のフォーマットは 44.1 kHz サンプリング周波数, 16bit RIFF WAV 形式です.

For speech recognition evaluation, test set lists for the subsets, test_{basic, hard, daily}.txt, are included in the corpus. The audio format is 44.1 kHz-sampled, 16-bit encoded RIFF WAV format.

サブセット / Subset カテゴリ / Category 説明 / Description 発話数 / #utts 時間長 / Dur. [h]
Basic - 通常音声と同じ文 / Parallel to control speech 4,701 6.89
Hard CF 聞き取りづらい音素を含む文 / Phonemic confusion 2,094 1.84
PCS 音声的に混同しやすい文 / Phonetically confusable 75 0.07
SUS 合文法無意味文 / Semantically unpredictable 73 0.09
Daily Everyday 家族や友達との会話 / Chat with family and friends 2,098 2.62
Research 研究者との議論 / Discussions with researchers 927 1.39
Work 公共スペースでの会話 / Conversation in public facilities 721 1.06
Emotion 感情表現 / Emotion expression and statements 414 0.54
Others 緊急,季節,趣味 / Urgency, seasons, or hobbies 399 0.63
合計 / Total - - 11,502 15.12

音声認識(ASR)ベンチマーク / ASR benchmark

本コーパスを使った音声認識実験の結果です.実験の詳細は文献を参照してください.単語誤り率(WER),BERTScore,Human eval. の値は,全てのテストセットリストの結果を平均したものです.

The table lists the results of our ASR experiments. See our paper for the details. Values are aggregated among test set lists for each of WER (word error rate), BERTScore, Human eval.

モデル / model WER BERTScore Human eval.
Whisper large v3 (no finetuning) 0.934 0.654 1.217
+ finetuning (13.0h) 0.283 0.915 3.748
+ finetuning + LLM 0.257 0.917 4.198

音声認識デモ / ASR demo

SS-JDSC ASR Demonstration

上記の画像をクリックするとYouTubeでデモ動画を再生します。

Click the image above to watch the demo video on YouTube.

ダウンロード / Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="JDSC-Project/SS-JDSC",
    repo_type="dataset",
    local_dir="SS-JDSC",
    local_dir_use_symlinks=False,
    resume_download=True,
    max_workers=8,
)

ロード / Load

from datasets import load_dataset, Audio, Features, Value

# 1. Define schema to ensure correct Audio decoding
ds_features = Features({
    "audio": Audio(sampling_rate=16000),
    "sentence": Value("string"),
    "subset": Value("string"),
})

# 2. Select Training Configuration
# Options: 
#   [Cumulative Mix] 
#   "train_0.4h", "train_0.8h", "train_1.6h", "train_3.2h", "train_6.5h", "train_13.0h"
#   [Domain Specific]
#   "train_basic-only_3.1h", "train_basic-only_6.2h"
#   "train_daily-only_2.8h", "train_daily-only_5.6h"
#   "train_hard-only_0.8h", "train_hard-only_1.6h"
TRAIN_CONFIG = "train_13.0h" 

# 3. Load Training Dataset
train_ds = load_dataset(
    "JDSC-Project/SS-JDSC", 
    name=TRAIN_CONFIG, 
    split="train", 
    features=ds_features
)

# 4. Load All Evaluation Datasets (Basic, Daily, Hard)
TEST_CONFIGS = ["test_basic", "test_daily", "test_hard"]
test_datasets = {
    cfg: load_dataset("JDSC-Project/SS-JDSC", name=cfg, split="train", features=ds_features)
    for cfg in TEST_CONFIGS
}

# Verification
print(f"Loaded Train set ({TRAIN_CONFIG}): {len(train_ds)} samples")
for name, ds in test_datasets.items():
    print(f"Loaded Test set ({name}): {len(ds)} samples")

メンバー / Member

  • Asahi Ogasawara
    Iwate University
  • Shinnosuke Takamichi
    Keio University / The University of Tokyo
  • Jianing Yang
    The University of Tokyo
  • Go Suenaga
    Independent Researcher
  • Tan Yiyu
    Iwate University

謝辞 / Acknowledgments

日本語

本コーパスの公開にあたり、多くの方々のご支援をいただきました。

高道准教授ならびに Yang 氏をはじめとする高道研究室の皆様には、本コーパスの核となる三つのサブセット構想の提案から、テキスト方針の策定、実験設計への助言、アノテーション、定期ミーティング、実験実施、論文執筆に至るまで、研究面で幅広いご支援をいただきました。

末永氏には、企画段階から継続的にご参加いただき、話者インタビュー、テキスト作成、アノテーション、論文指導を通して本プロジェクトを支えていただきました。

談准教授には、指導教官として本共同研究全体を通じて温かく力強いご指導をいただきました。

また、聞き取り評価に協力してくださった家族・友人・協力者の皆様にも心より感謝いたします。皆様の支えのおかげで、本コーパスは ICASSP 2026 において採択され、こうしてオープンに公開することができました。

本データセットが、構音障害音声研究および音声認識研究の発展に寄与し、コミュニティの中で活用・発展していくことを願っております。

(2026/02/12 小笠原 朝陽)


English

This dataset has been made possible thanks to the support of many contributors.

We would like to express our sincere gratitude to Associate Professor Takamichi and Yang, as well as the members of the Takamichi Laboratory, for their wide-ranging support. Their contributions include the initial proposal of the three core subset concepts, guidance on text design policy, advice on experimental design, annotation support, regular research discussions, experiment implementation, and assistance with manuscript preparation.

We are also deeply grateful to Mr. Suenaga for his continuous involvement from the project planning stage, including participation in meetings, speaker interviews, text creation, annotation, and guidance on the manuscript.

We would like to thank Associate Professor Dan for his dedicated supervision and consistent support throughout this collaborative research.

Finally, we sincerely thank the family members, friends, and collaborators who participated in the intelligibility evaluations. With their support, this corpus was accepted at ICASSP 2026 and is now publicly available.

We hope this dataset will contribute to further advances in dysarthric speech research and automatic speech recognition, and that it will continue to evolve through engagement within the research community.

Asahi Ogasawara
February 12, 2026

ライセンス / License

CC BY-NC-SA 4.0

引用 / Citation

このコーパスを研究で使用する場合は、必ず以下の論文の引用をお願いします。

If you use the SS-JDSC dataset in your research, please cite the following paper:

@inproceedings{ogasawara2026ssjdsc,
  title     = {{SS-JDSC}: Single-Speaker Japanese Dysarthric Speech Corpus},
  author    = {Ogasawara, Asahi and Takamichi, Shinnosuke and Yang, Jianing and Suenaga, Go and Tan, Yiyu},
  booktitle = {Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year      = {2026},
  note      = {to appear}
}

お問い合わせ / Contact

研究上のご質問、共同研究のご相談、または拡張的な利用
(商用利用、追加アノテーション、共同研究など)に関するお問い合わせは、
以下までご連絡ください。

For questions, collaborations, or requests for extended use cases
(e.g., commercial use, additional annotations, or joint research),
please contact the project lead:

Downloads last month
35