---
dataset_info:
  features:
  - name: surah_id
    dtype: int32
  - name: ayah_id
    dtype: int32
  - name: word_id
    dtype: string
  - name: word_index
    dtype: int32
  - name: word_ar
    dtype: string
  - name: word_en
    dtype: string
  - name: word_tr
    dtype: string
  - name: surah_name_ar
    dtype: string
  - name: surah_name_en
    dtype: string
  - name: ayah_ar
    dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 2164094570
    num_examples: 77429
  download_size: 2049220107
  dataset_size: 2164094570
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Quran-MD - Word Level

This dataset is part of the complete Quran-MD dataset available here: **[Complete Quran-MD](https://huggingface.co/datasets/Buraaq/quran-audio-text-dataset)**

## Dataset Description

This dataset contains **77,429 individual word-level audio recordings** from the complete Quran with precise pronunciations. Each sample includes a single word pronunciation along with its Arabic text, English meaning, transliteration, and contextual information from the source verse.


### Dataset Summary

- **📝 Audio Samples**: 77,429 MP3 files (individual word pronunciations)
- **📚 Coverage**: Complete Quran vocabulary (114 surahs)
- **🌐 Languages**: Arabic (original), English (meanings), Transliteration
- **🎧 Audio Quality**: High-quality word-level segmentation
- **⏱️ Duration**: Approximately ~20 hours of word-level audio
- **🔍 Granularity**: Individual word pronunciation with position context
- **Morphological Diversity**: Various grammatical forms and inflections


### Supported Tasks

- **🔤 Pronunciation Training**: Learn individual Arabic word pronunciation
- **📊 Linguistic Analysis**: Study Quranic vocabulary and morphology
- **🎯 Word Spotting**: Train word detection models
- **📖 Vocabulary Building**: Arabic language learning
- **🔍 Audio Segmentation**: Word-level audio alignment research
- **🎙️ Fine-grained ASR**: Word-level Arabic speech recognition


### Languages

- **Arabic** (Classical Arabic - Quranic Arabic)
- **English** (translations)
- **Transliteration** (romanized Arabic)


## Dataset Structure

Each sample contains individual word information with contextual metadata:

```python
{
    "surah_id": int32,           # Source surah (chapter) number
    "ayah_id": int32,            # Source ayah (verse) number  
    "word_id": string,           # Unique word identifier
    "word_index": int32,         # Position of word within the ayah (0-based)
    "word_ar": string,           # Individual word in Arabic
    "word_en": string,           # Word meaning/translation in English
    "word_tr": string,           # Word transliteration (romanized)
    "surah_name_ar": string,     # Context: Source surah name in Arabic
    "surah_name_en": string,     # Context: Source surah name in English
    "ayah_ar": string,           # Context: Complete ayah text for reference
    "audio": Audio(),            # Audio file (MP3 format) - isolated word
}
```

### Sample Data Example

```python
{
    "surah_id": 1,
    "ayah_id": 1,
    "word_id": "1:1:1",
    "word_index": 0,
    "word_ar": "بِسْمِ",
    "word_en": "In the name",
    "word_tr": "Bismi", 
    "surah_name_ar": "الفاتحة",
    "surah_name_en": "Al-Fatiha",
    "ayah_ar": "بِسْمِ اللَّهِ الرَّحْمَٰنِ الرَّحِيمِ",
    "audio": <Audio object>  # Isolated pronunciation of "بِسْمِ"
}
```


## Usage Examples

### Loading the Dataset

```python
from datasets import load_dataset

# Load the Word-level dataset  
dataset = load_dataset("Buraaq/quran-md-words")

print(f"Dataset size: {len(dataset):,} word recordings")
print(f"Sample keys: {list(dataset[0].keys())}")

# Access first sample
first_word = dataset[0]
print(f"Word: {first_word['word_ar']} ({first_word['word_tr']})")
print(f"Meaning: {first_word['word_en']}")
print(f"Context: {first_word['ayah_ar']}")
print(f"Position: Word {first_word['word_index'] + 1} in Surah {first_word['surah_id']}, Ayah {first_word['ayah_id']}")
```

### Play Word Audio

```python
import IPython.display as ipd

# Play individual word pronunciation
first_word = dataset[0]
print(f"Playing pronunciation of: {first_word['word_ar']} ({first_word['word_tr']})")
ipd.Audio(first_word["audio"]["array"], rate=first_word["audio"]["sampling_rate"])
```

### Vocabulary Analysis

```python
# Find all instances of a specific word
target_word = "اللَّهِ"  # "Allah"
allah_instances = dataset.filter(lambda x: x["word_ar"] == target_word)
print(f"Word '{target_word}' appears {len(allah_instances):,} times")

# Show contexts where this word appears
for i in range(min(5, len(allah_instances))):
    instance = allah_instances[i]
    print(f"Surah {instance['surah_id']}, Ayah {instance['ayah_id']}: {instance['ayah_ar']}")
```

### Word Length Analysis

```python
# Analyze word lengths
word_lengths = [len(sample["word_ar"]) for sample in dataset.select(range(1000))]
print(f"Average word length: {sum(word_lengths)/len(word_lengths):.2f} characters")
print(f"Shortest word: {min(word_lengths)} characters")  
print(f"Longest word: {max(word_lengths)} characters")

# Find shortest and longest words
shortest_words = dataset.filter(lambda x: len(x["word_ar"]) == min(word_lengths))
longest_words = dataset.filter(lambda x: len(x["word_ar"]) == max(word_lengths))

print(f"\\nShortest words examples:")
for i in range(min(3, len(shortest_words))):
    word = shortest_words[i]
    print(f"  {word['word_ar']} ({word['word_tr']}) - {word['word_en']}")
```

### Build Vocabulary Dictionary

```python
# Create vocabulary with meanings
vocabulary = {}
unique_words = set()

for sample in dataset.select(range(10000)):  # Sample for performance
    word_ar = sample["word_ar"]
    if word_ar not in vocabulary:
        vocabulary[word_ar] = {
            "transliteration": sample["word_tr"],
            "meanings": [sample["word_en"]],
            "count": 1,
            "contexts": [sample["ayah_ar"]]
        }
        unique_words.add(word_ar)
    else:
        vocabulary[word_ar]["count"] += 1
        if sample["word_en"] not in vocabulary[word_ar]["meanings"]:
            vocabulary[word_ar]["meanings"].append(sample["word_en"])

print(f"Unique words found: {len(unique_words):,}")
print(f"Most frequent words:")

# Sort by frequency
sorted_vocab = sorted(vocabulary.items(), key=lambda x: x[1]["count"], reverse=True)
for word, data in sorted_vocab[:10]:
    print(f"  {word} ({data['transliteration']}): {data['count']} times - {', '.join(data['meanings'])}")
```

### Filter by Word Position

```python
# Find all words that start verses (position 0)
verse_starters = dataset.filter(lambda x: x["word_index"] == 0)
print(f"Words that start verses: {len(verse_starters):,}")

# Find common verse-starting words
starter_words = {}
for sample in verse_starters.select(range(1000)):
    word = sample["word_ar"]
    starter_words[word] = starter_words.get(word, 0) + 1

print("Most common verse-starting words:")
for word, count in sorted(starter_words.items(), key=lambda x: x[1], reverse=True)[:10]:
    print(f"  {word}: {count} verses")
```

### Extract Words from Specific Ayah

```python
# Get all words from Al-Fatiha, first ayah
al_fatiha_words = dataset.filter(lambda x: x["surah_id"] == 1 and x["ayah_id"] == 1)
print(f"Al-Fatiha first ayah has {len(al_fatiha_words)} words")

# Display words in order
words_ordered = sorted(al_fatiha_words, key=lambda x: x["word_index"])
ayah_reconstruction = []

for word in words_ordered:
    print(f"Word {word['word_index'] + 1}: {word['word_ar']} ({word['word_tr']}) - {word['word_en']}")
    ayah_reconstruction.append(word['word_ar'])

print(f"\\nReconstructed ayah: {' '.join(ayah_reconstruction)}")
print(f"Original ayah: {words_ordered[0]['ayah_ar']}")
```


## Advanced Usage

### Building Pronunciation Models

```python
# Extract features for pronunciation modeling
import librosa
import numpy as np

def extract_word_features(audio_array, sample_rate):
    # Extract MFCC features
    mfccs = librosa.feature.mfcc(y=audio_array, sr=sample_rate, n_mfcc=13)
    # Extract spectral features
    spectral_centroids = librosa.feature.spectral_centroid(y=audio_array, sr=sample_rate)
    return {
        'mfccs': mfccs.mean(axis=1),
        'spectral_centroid': spectral_centroids.mean()
    }

# Process sample words
features_data = []
for i in range(100):  # Sample processing
    sample = dataset[i]
    audio = sample["audio"]["array"]
    sr = sample["audio"]["sampling_rate"]
    features = extract_word_features(audio, sr)
    features_data.append({
        'word': sample["word_ar"],
        'features': features
    })
```

### Word Similarity Analysis

```python
# Find similar sounding words
from difflib import SequenceMatcher

def find_similar_words(target_word, dataset_sample, threshold=0.6):
    similar_words = []
    target_tr = None
    
    # Find target word's transliteration
    for sample in dataset_sample:
        if sample["word_ar"] == target_word:
            target_tr = sample["word_tr"]
            break
    
    if not target_tr:
        return similar_words
    
    # Find similar transliterations
    for sample in dataset_sample:
        similarity = SequenceMatcher(None, target_tr.lower(), sample["word_tr"].lower()).ratio()
        if similarity > threshold and sample["word_ar"] != target_word:
            similar_words.append({
                'word': sample["word_ar"],
                'transliteration': sample["word_tr"],
                'meaning': sample["word_en"],
                'similarity': similarity
            })
    
    return sorted(similar_words, key=lambda x: x['similarity'], reverse=True)

# Example usage
similar = find_similar_words("اللَّهِ", dataset.select(range(1000)))
print("Words similar to 'اللَّهِ':")
for word_info in similar[:5]:
    print(f"  {word_info['word']} ({word_info['transliteration']}) - {word_info['similarity']:.2f}")
```

## Citation

```bibtex
@inproceedings{salman2025quran,
  title={Quran-MD: A Fine-Grained Multimodal Dataset of the Quran},
  author={Salman, Muhammad Umar and Qazi, Mohammad Areeb and Alam, Mohammed Talha},
  booktitle={5th Muslims in ML Workshop co-located with NeurIPS 2025}
}
```

## Related Datasets

- **📝 [Quran-MD - Ayah](https://huggingface.co/datasets/Buraaq/quran-md-ayahs)**: Complete verse recitations (187,080 samples)
- **🏠 [Complete Quran-MD Overview](https://huggingface.co/datasets/Buraaq/quran-audio-text-dataset)**: Full dataset collection documentation

## Contact

- **Hugging Face Profile**: umarsalman
- **GitHub Profile**: umar1997
- **Dataset Issues**: Please open an issue on the repository.

Feel free to reach out for dataset issues, research collaborations, or community contributions.

---

*"And We have certainly made the Quran easy for remembrance, so is there any who will remember?"* - Quran 54:17