RetaSy/quranic_audio_dataset
收藏资源简介：
我们探索了众包一个精心标注的古兰经数据集的可能性，在此基础上可以构建AI模型以简化学习过程。特别是，我们使用基于志愿者的众包类型，并实现了一个众包API来收集音频资源。我们开发了一个名为Quran Voice的众包平台来标注收集的音频资源。结果，我们从11个以上非阿拉伯国家的1287名参与者中收集了约7000个古兰经诵读音频，并从数据集中标注了1166个诵读音频，分为六个类别。我们达到了0.77的众包准确率，标注者之间的评分一致性为0.63，算法分配的标签与专家判断之间的一致性为0.89。
We explored the feasibility of crowdsourcing a meticulously annotated Quranic dataset, upon which AI models can be built to streamline the learning process. Specifically, we adopted a volunteer-based crowdsourcing approach and implemented a crowdsourcing API for audio resource collection. We developed a crowdsourcing platform named Quran Voice to annotate the collected audio samples. As a result, we collected approximately 7,000 Quran recitation audio samples from 1,287 participants across more than 11 non-Arab countries, and annotated 1,166 recitation audio samples from the dataset, which were categorized into six classes. We achieved a crowdsourcing accuracy of 0.77, an inter-annotator agreement score of 0.63, and a consistency of 0.89 between algorithm-assigned labels and expert judgments.
数据集概述
数据集名称
- 名称: Quranic Audio Dataset - Crowdsourced and Labeled Recitation from Non-Arabic Speakers
数据集特征
- audio (
dict): 包含音频文件路径、解码后的音频数组和采样率。 - Surah (
string): 古兰经的章节。 - Aya (
string): 古兰经章节中的特定诗句。 - duration_ms (
int64): 音频记录的持续时间，以毫秒为单位。 - create_date (
string): 音频记录的创建日期和时间。 - golden (
bool): 音频样本是否由专家标记。 - final_label (
string): 音频样本的共识标签，指示朗诵的分类。 - reciter_id (
string): 朗诵者的唯一标识符。 - reciter_country (
string): 朗诵者的国家。 - reciter_gender (
string): 朗诵者的性别。 - reciter_age (
string): 朗诵者的年龄。 - reciter_qiraah (
string): 朗诵者使用的朗诵风格。 - judgments_num (
int64): 每个音频样本的评判或注释数量。 - annotation_metadata (
string): 与每个音频样本的注释相关的元数据。
数据集结构
- 训练集 (
train): 包含6828个样本，总大小为1290351809.656字节。
任务类别
- automatic-speech-recognition
- audio-classification
语言
- ar (阿拉伯语)
标签
- Crowdsourcing
- Quranic recitation
- Non-Arabic Speakers
引用信息
@inproceedings{quran_audio_dataset:2024, author = {Raghad Salameh, Mohamad Al Mdfaa, Nursultan Askarbekuly, Manuel Mazzara}, title = {Quranic Audio Dataset: Crowdsourced and Labeled Recitation from Non-Arabic Speakers}, year = 2024, eprint = {2405.02675}, eprinttype = {arxiv}, eprintclass = {cs.SD}, url = {https://arxiv.org/abs/2405.02675}, language = {english} }
