hetchyy/quranic-universal-ayahs

---
license: apache-2.0
task_categories:
- automatic-speech-recognition
language:
- ar
tags:
- quran
- recitation
- forced-alignment
- word-timestamps
- audio-segmentation
- speech-recognition
- asr
- vad
- phoneme
version: v0.2.2
pretty_name: Qur'anic Universal Ayahs
size_categories:
- 1K<n<10K
configs:
- config_name: hafs_an_asim
data_files:
- split: minshawy_murattal
path: hafs_an_asim/minshawy_murattal-*
- split: ali_jaber
path: hafs_an_asim/ali_jaber-*
dataset_info:
config_name: hafs_an_asim
features:
- name: audio
dtype: audio
- name: surah
dtype: int32
- name: ayah
dtype: int32
- name: text
dtype: string
- name: segments
sequence:
sequence: int32
- name: words
sequence:
sequence: int32
splits:
- name: minshawy_murattal
num_bytes: 1571705367
num_examples: 6236
- name: ali_jaber
num_bytes: 0
num_examples: 6236
download_size: 1569809077
dataset_size: 1571705367
---

<p align="center">
<a href="https://huggingface.co/spaces/hetchyy/Quran-multi-aligner"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Demo-Qur'an%20Multi--Aligner-yellow" alt="Demo - Qur'an Multi-Aligner"></a>
<a href="https://huggingface.co/spaces/hetchyy/Quran-reciter-requests"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Request-New%20Reciter-ff69b4" alt="Request - New Reciter"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/releases/latest"><img src="https://img.shields.io/github/v/release/Wider-Community/quranic-universal-audio?label=Release" alt="Latest Release"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/blob/main/data/RECITERS.md"><img src="https://img.shields.io/badge/Reciters-337%20Available%20%7C%202%20Aligned-green" alt="Reciters"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/blob/main/data/RECITERS.md"><img src="https://img.shields.io/badge/Riwayat-14%20%2F%2020-green" alt="Riwayat"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache%202.0-orange" alt="License"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio"><img src="https://img.shields.io/github/stars/Wider-Community/quranic-universal-audio?style=social" alt="GitHub stars"></a>
</p>

# Qur'anic Universal Ayahs

Word-level aligned Qur'an recitation audio with precise timestamps derived from phoneme-level forced alignment.

## Dataset Description

Each row is one verse (ayah) of the Qur'an, with:
- **Audio clip** of the verse recitation, trimmed to speech boundaries
- **Word-level timestamps** in milliseconds, relative to the audio clip
- **Pause-based segments** showing how the recitation was naturally divided by silences
- **Arabic text** from alignment matching (reflects what was actually recited, including any repetitions)

## Usage

```python
from datasets import load_dataset

# Load a specific reciter (subset = riwayah, split = reciter)
ds = load_dataset("hetchyy/quranic-universal-ayahs", "hafs_an_asim", split="minshawy_murattal")

# Access a verse
verse = ds[0]
print(verse["surah"], verse["ayah"]) # 1 1
print(verse["text"]) # Arabic text
print(verse["words"]) # [[1, 0, 400], [2, 400, 800], ...]

# Play audio (in a notebook)
from IPython.display import Audio
Audio(verse["audio"]["array"], rate=verse["audio"]["sampling_rate"])
```
## Schema

| Column | Type | Description |
|--------|------|-------------|
| `audio` | `Audio` | Verse audio clip, trimmed to speech boundaries |
| `hafs_an_asim` | `ali_jaber` | Ali Jaber | 6,236 | everyayah.com |
| `surah` | `int32` | Surah number (1-114) |
| `ayah` | `int32` | Verse number within surah |
| `text` | `string` | Arabic text of the verse from alignment |
| `segments` | `[[int, int, int, int]]` | Pause-based segments (ms, relative to clip) |
| `words` | `[[int, int, int]]` | Word-level timestamps (ms, relative to clip) |

### Column Details

**`segments`** — Each segment is `[word_from, word_to, start_ms, end_ms]`. Represents a continuous speech region between pauses. Word indices are 1-based.

**`words`** — Each word is `[word_index, start_ms, end_ms]`. Word-level timestamps from phoneme-level forced alignment (MFA). Word indices are 1-based.

## Configs

Subset (config) is the riwayah, split is the reciter.

| Subset | Split | Reciter | Verses | Audio Source |
|--------|-------|---------|--------|-------------|
| `hafs_an_asim` | `minshawy_murattal` | Mohamed Siddiq El-Minshawi (Murattal) | 6,236 | everyayah.com |

## Pipeline

Audio is processed through a multi-stage pipeline:
1. **VAD segmentation** — Detect speech regions using a recitation-specific VAD model
2. **Phoneme-level ASR** — CTC-based recognition with wav2vec2
3. **Dynamic programming alignment** — Match recognized phonemes against known Qur'anic reference text
4. **MFA forced alignment** — Montreal Forced Aligner produces phoneme-level timestamps, from which word boundaries are derived

## Notes

- All timestamps are in **milliseconds**, relative to the start of the audio clip
- Word indices are **1-based**
- Word timestamps are padded forward within each segment so there are no gaps between consecutive words. Gaps only occur across segment boundaries (pauses in recitation).
- Text is derived from segment alignment and preserves any repetitions in the recitation
- Audio clips are trimmed to the first/last word boundaries (silence before/after is removed)

## License

[Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)

license: apache-2.0
task_categories:
- 自动语音识别（Automatic Speech Recognition）
language:
- ar（阿拉伯语）
tags:
- 古兰经（Quran）
- 诵读（recitation）
- 强制对齐（forced-alignment）
- 词级时间戳（word-timestamps）
- 音频分割（audio-segmentation）
- 语音识别（speech-recognition）
- ASR（Automatic Speech Recognition）
- VAD（Voice Activity Detection，语音活动检测）
- 音素（phoneme）
version: v0.2.2
pretty_name: 古兰经通用经文（Qur'anic Universal Ayahs）
size_categories:
- 1K<n<10K
configs:
- config_name: hafs_an_asim
data_files:
- split: minshawy_murattal
path: hafs_an_asim/minshawy_murattal-*
- split: ali_jaber
path: hafs_an_asim/ali_jaber-*
dataset_info:
config_name: hafs_an_asim
features:
- name: audio（音频）
dtype: audio
- name: surah（章号）
dtype: int32
- name: ayah（经文节数）
dtype: int32
- name: text（文本）
dtype: string
- name: segments（分段）
sequence:
sequence: int32
- name: words（词）
sequence:
sequence: int32
splits:
- name: minshawy_murattal
num_bytes: 1571705367
num_examples: 6236
- name: ali_jaber
num_bytes: 0
num_examples: 6236
download_size: 1569809077
dataset_size: 1571705367

<p align="center">
<a href="https://huggingface.co/spaces/hetchyy/Quran-multi-aligner"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Demo-Qur'an%20Multi--Aligner-yellow" alt="Demo - 古兰经多对齐工具"></a>
<a href="https://huggingface.co/spaces/hetchyy/Quran-reciter-requests"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Request-New%20Reciter-ff69b4" alt="Request - 申请新增诵读家"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/releases/latest"><img src="https://img.shields.io/github/v/release/Wider-Community/quranic-universal-audio?label=Release" alt="Latest Release - 最新版本"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/blob/main/data/RECITERS.md"><img src="https://img.shields.io/badge/Reciters-337%20Available%20%7C%202%20Aligned-green" alt="Reciters - 可用诵读家337位 | 已对齐2位"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/blob/main/data/RECITERS.md"><img src="https://img.shields.io/badge/Riwayat-14%20%2F%2020-green" alt="Riwayat - 传述体系14/20"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache%202.0-orange" alt="License - 许可证"></a>
<a href="https://github.com/Wider-Community/quranic-universal-audio"><img src="https://img.shields.io/github/stars/Wider-Community/quranic-universal-audio?style=social" alt="GitHub stars - GitHub星标"></a>
</p>

# 古兰经通用经文数据集（Qur'anic Universal Ayahs）

本数据集包含经词级对齐的古兰经诵读音频，其精确时间戳由音素级强制对齐生成。

## 数据集说明

每一行对应古兰经的一节经文（ayah），包含以下内容：
- **音频片段**：该经文诵读的音频，已裁剪至语音起止边界
- **词级时间戳**：以毫秒为单位，相对于音频片段的起始点
- **基于停顿的分段信息**：展示诵读过程中由自然停顿划分的连续语音区域
- **阿拉伯语文本**：来自对齐匹配的结果，可反映实际诵读内容，包括重复部分

## 使用方法

python
from datasets import load_dataset

# 加载指定诵读家的子集（配置项对应传述体系，拆分集对应诵读家）
ds = load_dataset("hetchyy/quranic-universal-ayahs", "hafs_an_asim", split="minshawy_murattal")

# 访问某一节经文
verse = ds[0]
print(verse["surah"], verse["ayah"]) # 输出示例：1 1（章数1，经文节数1）
print(verse["text"]) # 阿拉伯语文本
print(verse["words"]) # 格式：[[1, 0, 400], [2, 400, 800], ...]

# 在Notebook中播放音频
from IPython.display import Audio
Audio(verse["audio"]["array"], rate=verse["audio"]["sampling_rate"])

## 数据结构

| 列名 | 数据类型 | 描述 |
|--------|------|-------------|
| `audio` | `Audio` | 经文音频片段，已裁剪至语音起止边界 |
| `surah` | `int32` | 章号（1-114） |
| `ayah` | `int32` | 该章内的经文节数 |
| `text` | `string` | 来自对齐匹配的经文章节阿拉伯语文本 |
| `segments` | `[[int, int, int, int]]` | 基于停顿的分段信息（毫秒，相对于音频片段） |
| `words` | `[[int, int, int]]` | 词级时间戳信息（毫秒，相对于音频片段） |

### 列详情

**`segments`（分段信息）**：每个分段的格式为`[起始词索引, 结束词索引, 起始毫秒数, 结束毫秒数]`，代表一次停顿间的连续语音区域。词索引采用1-based（从1开始计数）。

**`words`（词信息）**：每个词的格式为`[词索引, 起始毫秒数, 结束毫秒数]`，时间戳由音素级强制对齐（Montreal Forced Aligner，MFA）生成。词索引采用1-based（从1开始计数）。

## 配置项说明

配置项（subset）对应古兰经诵读的传述体系（riwayah），拆分集（split）对应诵读家。

| 传述体系 | 诵读家拆分集 | 诵读家 | 经文总数 | 音频来源 |
|--------|-------|---------|--------|-------------|
| `hafs_an_asim` | `minshawy_murattal` | 穆罕默德·西迪克·埃尔·明沙维（穆拉塔尔风格诵读） | 6236 | everyayah.com |

## 数据处理流程

音频将通过以下多阶段流程处理：
1. **语音活动检测（VAD）分段**：使用针对诵读场景优化的VAD模型识别语音区域
2. **音素级自动语音识别（ASR）**：基于wav2vec2的CTC（联结主义时间分类）识别
3. **动态规划对齐**：将识别出的音素与已知的古兰经参考文本进行匹配
4. **MFA强制对齐**：通过蒙特利尔强制对齐工具（Montreal Forced Aligner，MFA）生成音素级时间戳，并由此推导词边界信息

## 注意事项

- 所有时间戳单位均为**毫秒**，相对于音频片段的起始点
- 词索引采用**1-based（从1开始计数）**
- 同一分段内的词时间戳会向前补全，确保连续词之间无间隙；仅在分段边界（即诵读停顿处）存在间隙
- 文本来自分段对齐结果，可保留诵读中的所有重复内容
- 音频片段已裁剪至首个和末次词的边界（移除前后的静音片段）

## 许可证

[Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)