viewer: true
dataset_info:
- config_name: Chinese
features:
- name: audio
dtype:
audio:
sampling_rate: 16000
- name: text
dtype: string
- name: duration
dtype: float64
splits:
- name: train
num_bytes: 182566135.142
num_examples: 1242
- name: eval
num_bytes: 12333509
num_examples: 91
- name: test
num_bytes: 33014034
num_examples: 225
download_size: 227567289
dataset_size: 227913678.142
- config_name: English
features:
- name: audio
dtype:
audio:
sampling_rate: 16000
- name: text
dtype: string
- name: duration
dtype: float64
splits:
- name: train
num_bytes: 2789314997.152
num_examples: 25512
- name: eval
num_bytes: 299242087.632
num_examples: 2816
- name: test
num_bytes: 553873172.749
num_examples: 4751
download_size: 3627859275
dataset_size: 3642430257.533
- config_name: French
features:
- name: audio
dtype:
audio:
sampling_rate: 16000
- name: text
dtype: string
- name: duration
dtype: float64
splits:
- name: train
num_bytes: 168642145.231
num_examples: 1403
- name: eval
num_bytes: 5164908
num_examples: 42
- name: test
num_bytes: 42780388
num_examples: 344
download_size: 216118671
dataset_size: 216587441.231
- config_name: German
features:
- name: audio
dtype: audio
- name: text
dtype: string
- name: duration
dtype: float64
splits:
- name: train
num_bytes: 181312217.029
num_examples: 1443
- name: test
num_bytes: 137762006.256
num_examples: 1091
- name: eval
num_bytes: 35475098
num_examples: 287
download_size: 354494147
dataset_size: 354549321.285
- config_name: Vietnamese
features:
- name: audio
dtype: audio
- name: text
dtype: string
- name: duration
dtype: float64
splits:
- name: train
num_bytes: 56584901.453
num_examples: 2773
- name: test
num_bytes: 69598082.31
num_examples: 3437
- name: dev
num_bytes: 57617298.896
num_examples: 2912
download_size: 181789393
dataset_size: 183800282.659
configs:
- config_name: Chinese
data_files:
- split: train
path: Chinese/train-*
- split: eval
path: Chinese/eval-*
- split: test
path: Chinese/test-*
- config_name: English
data_files:
- split: train
path: English/train-*
- split: eval
path: English/eval-*
- split: test
path: English/test-*
- config_name: French
data_files:
- split: train
path: French/train-*
- split: eval
path: French/eval-*
- split: test
path: French/test-*
- config_name: German
data_files:
- split: train
path: German/train-*
- split: test
path: German/test-*
- split: eval
path: German/eval-*
- config_name: Vietnamese
data_files:
- split: train
path: Vietnamese/train-*
- split: test
path: Vietnamese/test-*
- split: dev
path: Vietnamese/dev-*
MultiMed: Multilingual Medical Speech Recognition via Attention Encoder Decoder
Description:
Multilingual automatic speech recognition (ASR) in the medical domain serves as a foundational task for various downstream applications such as speech translation, spoken language understanding, and voice-activated assistants. This technology enhances patient care by enabling efficient communication across language barriers, alleviating specialized workforce shortages, and facilitating improved diagnosis and treatment, particularly during pandemics. In this work, we introduce \textit{MultiMed}, a collection of small-to-large end-to-end ASR models for the medical domain, spanning five languages: Vietnamese, English, German, French, and Mandarin Chinese, together with the corresponding real-world ASR dataset. To our best knowledge, \textit{MultiMed} stands as the largest and the first multilingual medical ASR dataset, in terms of total duration, number of speakers, diversity of diseases, recording conditions, speaker roles, unique medical terms, accents, and ICD-10 codes.
Please cite this paper: https://arxiv.org/abs/2404.05659
@inproceedings{VietMed_dataset,
title={VietMed: A Dataset and Benchmark for Automatic Speech Recognition of Vietnamese in the Medical Domain},
author={Khai Le-Duc},
year={2024},
booktitle = {Proceedings of the Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
}
To load labeled data, please refer to our HuggingFace, Paperswithcodes.
For full dataset (labeled data + unlabeled data) and pre-trained models, please refer to Google Drive
Limitations:
Since this dataset is human-labeled, 1-2 ending/starting words present in the recording might not be present in the transcript. That's the nature of human-labeled dataset, in which humans can't distinguish words that are faster than 1 second. In contrast, forced alignment could solve this problem because machines can "listen" words in 10ms-20ms. However, forced alignment only learns what it is taught by humans. Therefore, no transcript is perfect. We will conduct human-machine collaboration to get "more perfect" transcript in the next paper.
Contact:
If any links are broken, please contact me for fixing!
Thanks Phan Phuc for dataset viewer <3
Le Duc Khai
University of Toronto, Canada
Email: [email protected]
GitHub: https://github.com/leduckhai