README.md

---
license: cc-by-4.0
task_categories:
- text-to-speech
- automatic-speech-recognition
- audio-to-audio
language:
- ur
tags:
- Urdu
- TTS
- LargeScaleDataset
pretty_name: Munch
---
# Munch - Large-Scale Urdu Text-to-Speech Dataset

[![Dataset](https://img.shields.io/badge/🤗%20Dataset-Munch-blue)](https://huggingface.co/datasets/humair025/Munch)
[![Hashed Index](https://img.shields.io/badge/🤗%20Index-hashed__data-green)](https://huggingface.co/datasets/humair025/hashed_data)
[![Size](https://img.shields.io/badge/Size-1.27TB-orange)]()
[![Rows](https://img.shields.io/badge/Rows-4.17M-brightgreen)]()
[![License](https://img.shields.io/badge/License-CC--BY--4.0-yellow)]()

## 📖 Dataset Description

**Munch** is a large-scale Urdu Text-to-Speech (TTS) dataset containing high-quality audio recordings paired with Urdu text transcripts. The dataset features multiple voice variations and natural pronunciation patterns suitable for training and evaluating Urdu TTS models.




### Rough Assumption :

    4.17 million audio clips, if each 20 seconds long, total about 11,500+ hours of audio.


### Key Features

- 🎤 **13 Different Voices**: alloy, echo, fable, onyx, nova, shimmer, coral, verse, ballad, ash, sage, amuch, dan
- 🗣️ **Natural Urdu Pronunciation**: Proper handling of Urdu script, punctuation, and intonation
- 📊 **Large Scale**: 4,167,500 audio-text pairs
- 🎵 **High Quality Audio**: PCM16 format, 22.05 kHz sample rate
- 💾 **Efficient Storage**: Parquet format with compression
- 📇 **Lightweight Index Available**: [Hashed index](https://huggingface.co/datasets/humair025/hashed_data) for exploration without downloading full dataset

### Dataset Statistics

| Metric | Value |
|--------|-------|
| Total Size | 1.27 TB |
| Total Rows | 4,167,500 |
| Number of Files | ~8,300 parquet files |
| Audio Format | PCM16 (raw audio bytes) |
| Sample Rate | 22,050 Hz |
| Bit Depth | 16-bit signed integer |
| Text Language | Urdu (with occasional mixed language) |
| Voice Count | 13 unique voices |
| Avg Audio Size | ~50 KB to 5MB per sample |
| Avg Duration | ~3-5 seconds per sample |
| Total Duration | ~7,500-15,800 hours of audio |

### 🔗 Companion Dataset

For efficient exploration without downloading the full 1.27 TB dataset, use the [**Munch Hashed Index**](https://huggingface.co/datasets/humair025/hashed_data):
- 📊 Contains all metadata + SHA-256 hashes of audio
- 💾 Only ~1 GB (99.92% smaller)
- ⚡ Search 4.17M records in seconds
- 🎯 Selectively download only what you need

### Related Datasets

- **This Dataset (v1)**: [humair025/Munch](https://huggingface.co/datasets/humair025/Munch) - 1.27 TB, 4.17M samples
- **Munch-1 (v2)**: [humair025/munch-1](https://huggingface.co/datasets/humair025/munch-1) - 3.28 TB, 3.86M samples (newer version)
- **Hashed Index (v1)**: [humair025/hashed_data](https://huggingface.co/datasets/humair025/hashed_data) - Index for this dataset
- **Hashed Index (v2)**: [humair025/hashed_data_munch_1](https://huggingface.co/datasets/humair025/hashed_data_munch_1) - Index for Munch-1

---

## 🚀 Quick Start

### Installation

```bash
pip install datasets pandas numpy scipy
```

### Basic Usage

```python
from datasets import load_dataset
import numpy as np
import io
from scipy.io import wavfile
import IPython.display as ipd

# Load a specific file
ds = load_dataset(
    "humair025/Munch",
    data_files="tts_data_20251203_125841_0a26c418.parquet",
    split="train"
)

# Helper function to convert PCM16 bytes to WAV
def pcm16_bytes_to_wav(pcm_bytes, sample_rate=22050):
    audio_array = np.frombuffer(pcm_bytes, dtype=np.int16)
    wav_io = io.BytesIO()
    wavfile.write(wav_io, sample_rate, audio_array)
    wav_io.seek(0)
    return wav_io

# Play first audio sample
row = ds[0]
wav_io = pcm16_bytes_to_wav(row['audio_bytes'])
ipd.display(ipd.Audio(wav_io, rate=22050))

print(f"Text: {row['text']}")
print(f"Voice: {row['voice']}")
```

### Efficient Exploration (Recommended)

Instead of downloading the full 1.27 TB dataset, start with the hashed index:

```python
from datasets import load_dataset
import pandas as pd

# Load the lightweight index (~1 GB)
index_ds = load_dataset("humair025/hashed_data", split="train")
index_df = pd.DataFrame(index_ds)

# Explore the dataset
print(f"Total samples: {len(index_df)}")
print(f"Voices: {index_df['voice'].unique()}")
print(f"Voice distribution:\n{index_df['voice'].value_counts()}")

# Find specific samples
ash_samples = index_df[index_df['voice'] == 'ash']
short_audio = index_df[index_df['audio_size_bytes'] < 40000]

# Download only what you need
files_needed = ash_samples['parquet_file_name'].unique()[:10]
ds = load_dataset(
    "humair025/Munch",
    data_files=list(files_needed),
    split="train"
)
```

### Load Multiple Files

```python
# Load first 10 files
ds = load_dataset(
    "humair025/Munch",
    data_files="tts_data_20251203_*.parquet",  # Wildcard pattern
    split="train"
)

print(f"Total samples: {len(ds)}")
```

### Batch Processing

```python
from huggingface_hub import HfApi

# Get all parquet files
api = HfApi()
files = api.list_repo_files(repo_id="humair025/Munch", repo_type="dataset")
parquet_files = [f for f in files if f.endswith('.parquet')]

print(f"Total files: {len(parquet_files)}")

# Load first 20 files
batch = parquet_files[:20]
ds = load_dataset(
    "humair025/Munch",
    data_files=batch,
    split="train"
)
```

---

## 📊 Dataset Structure

### Data Fields

Each row in the dataset contains:

| Field | Type | Description |
|-------|------|-------------|
| `id` | int | Paragraph ID (sequential) |
| `text` | string | Original Urdu text |
| `transcript` | string | TTS transcript (may differ slightly from input) |
| `voice` | string | Voice name used (e.g., "ash", "sage", "coral") |
| `audio_bytes` | bytes | Raw PCM16 audio data |
| `timestamp` | string | ISO format timestamp of generation (nullable) |
| `error` | string | Error message if generation failed (nullable) |

### Example Row

```python
{
    'id': 42,
    'text': 'یہ ایک نمونہ متن ہے۔',
    'transcript': 'یہ ایک نمونہ متن ہے۔',
    'voice': 'ash',
    'audio_bytes': b'\x00\x01...',  # PCM16 bytes
    'timestamp': '2025-12-03T13:03:14.123456',
    'error': None
}
```

---

## 🎯 Use Cases

### 1. **TTS Model Training**
Train Urdu text-to-speech models with diverse voice samples:
- Fine-tune existing TTS models
- Train voice cloning systems
- Develop multi-speaker TTS
- Create voice conversion models

### 2. **Speech Recognition**
Develop Urdu ASR systems:
- Train speech-to-text models
- Evaluate transcription accuracy
- Research Urdu phonetics
- Build pronunciation dictionaries

### 3. **Voice Research**
Study voice characteristics and patterns:
- Analyze voice similarity
- Research pronunciation patterns
- Study Urdu phonetics and prosody
- Compare voice quality metrics

### 4. **Audio Processing**
Develop audio processing pipelines:
- Audio enhancement
- Noise reduction
- Speech synthesis evaluation
- Audio quality assessment

### 5. **Linguistic Analysis**
Explore linguistic patterns:
- Text analysis and corpus linguistics
- Punctuation usage patterns
- Sentence structure analysis
- Code-switching research (Urdu-English)

---

## 🔧 Advanced Usage

### Voice Distribution Analysis

```python
import pandas as pd
from collections import Counter

# Using the hashed index (recommended)
index_ds = load_dataset("humair025/hashed_data", split="train")
index_df = pd.DataFrame(index_ds)

# Count voice usage
voice_counts = index_df['voice'].value_counts()
print("Voice Distribution:")
for voice, count in voice_counts.items():
    percentage = (count / len(index_df)) * 100
    print(f"  {voice}: {count:,} samples ({percentage:.2f}%)")
```

### Audio Length Analysis

```python
# Using the hashed index
avg_size = index_df['audio_size_bytes'].mean()
avg_duration = (avg_size / 2) / 22050  # bytes to seconds

print(f"Average audio size: {avg_size/1024:.2f} KB")
print(f"Average duration: {avg_duration:.2f} seconds")

# Duration distribution
durations = (index_df['audio_size_bytes'] / 2) / 22050
print(f"Min duration: {durations.min():.2f}s")
print(f"Max duration: {durations.max():.2f}s")
print(f"Median duration: {durations.median():.2f}s")
```

### Text Statistics

```python
# Text length analysis
text_lengths = index_df['text'].str.len()
word_counts = index_df['text'].str.split().str.len()

print(f"Average characters: {text_lengths.mean():.0f}")
print(f"Average words: {word_counts.mean():.0f}")
print(f"Longest text: {text_lengths.max()} characters")
```

### Duplicate Detection

```python
# Find duplicate audio using hashes
duplicates = index_df[index_df.duplicated(subset=['audio_bytes_hash'], keep=False)]

if len(duplicates) > 0:
    print(f"Found {len(duplicates):,} duplicate rows")
    print(f"Unique audio: {index_df['audio_bytes_hash'].nunique():,}")
    redundancy = (1 - index_df['audio_bytes_hash'].nunique()/len(index_df)) * 100
    print(f"Redundancy: {redundancy:.2f}%")
else:
    print("No duplicates found!")
```

### Export to WAV Files

```python
import os
from tqdm import tqdm

# Load specific samples
ds = load_dataset(
    "humair025/Munch",
    data_files="tts_data_20251203_*.parquet",
    split="train"
)

os.makedirs("audio_files", exist_ok=True)

for i, row in enumerate(tqdm(ds[:100])):  # First 100 samples
    wav_io = pcm16_bytes_to_wav(row['audio_bytes'])
    filename = f"audio_files/sample_{i:04d}_{row['voice']}.wav"
    with open(filename, 'wb') as f:
        f.write(wav_io.read())
```

### Selective Download by Voice

```python
# Using hashed index to find files
voice_of_interest = 'ash'
ash_files = index_df[index_df['voice'] == voice_of_interest]['parquet_file_name'].unique()

print(f"Files containing '{voice_of_interest}' voice: {len(ash_files)}")

# Download first 10 files with ash voice
ds = load_dataset(
    "humair025/Munch",
    data_files=list(ash_files[:10]),
    split="train"
)

print(f"Loaded {len(ds)} samples")
```

---

## 📝 Dataset Creation

This dataset was generated using a high-performance parallel TTS pipeline with the following characteristics:

### Generation Pipeline

- **Concurrent Processing**: 10-20 parallel workers
- **Voice Rotation**: Sequential rotation through 13 voices
- **Quality Control**: Automatic retry with exponential backoff
- **Fault Tolerance**: Checkpoint-based resumption
- **Smart Batching**: Efficient 500-row batches
- **API**: OpenAI-compatible TTS endpoints

### Pipeline Features

- ✅ Natural Urdu pronunciation with proper intonation
- ✅ Punctuation-aware pausing:
  - `؟` (question mark): 400ms pause with higher pitch
  - `!` (exclamation): 300ms pause with emphasis
  - `،` (comma): 500ms pause
  - `۔` (full stop): 1000ms pause
- ✅ Mixed-language support for technical terms
- ✅ Variable pacing for natural flow
- ✅ Error handling and logging

---

## ⚠️ Important Notes

### Audio Format
- Audio is stored as **raw PCM16 bytes** (not WAV files)
- Must be converted before playback (see examples above)
- Sample rate: 22,050 Hz
- Bit depth: 16-bit signed integer
- Channels: Mono (1 channel)

### Large Dataset Considerations
- 💾 **Size**: 1.27 TB total - download selectively
- 📦 **Files**: ~8,300 individual parquet files
- ⚡ **Streaming**: Recommended for full dataset access
- 🔄 **Batching**: Load files in batches to manage memory
- 📊 **Index First**: Use [hashed index](https://huggingface.co/datasets/humair025/hashed_data) to explore before downloading

### Recommended Workflow

1. **Explore**: Load the [hashed index](https://huggingface.co/datasets/humair025/hashed_data) (~1 GB)
2. **Filter**: Find samples matching your criteria
3. **Download**: Selectively download only needed parquet files
4. **Process**: Work with manageable subsets

### Potential Data Issues

⚠️ **Duplicates**: This dataset may contain duplicate audio samples. Use the hashed index for deduplication:

```python
# Get unique samples only
unique_df = index_df.drop_duplicates(subset=['audio_bytes_hash'], keep='first')
unique_files = unique_df['parquet_file_name'].unique()
```

⚠️ **Quality Variance**: Some samples may have:
- Low volume or clipping
- Mispronunciations (especially for rare words)
- Background noise
- Transcription differences from input text

---

## 📊 Performance Tips

### Memory Management

```python
# DON'T: Load entire dataset at once
# ds = load_dataset("humair025/Munch", split="train")  # 1.27 TB!

# DO: Use streaming mode
ds = load_dataset(
    "humair025/Munch",
    data_files="tts_data_20251203_*.parquet",
    split="train",
    streaming=True  # Stream data instead of loading all
)

# Process in batches
for i, batch in enumerate(ds.iter(batch_size=100)):
    # Process 100 samples at a time
    if i >= 10:  # Process only first 1000 samples
        break
```

### Efficient File Selection

```python
# Select specific date range
ds = load_dataset(
    "humair025/Munch",
    data_files="tts_data_20251203_*.parquet",  # Only Dec 3rd files
    split="train"
)

# Or specific time range
ds = load_dataset(
    "humair025/Munch",
    data_files="tts_data_20251203_1303*.parquet",  # Around 13:03
    split="train"
)

# Or use the index to find specific files
target_files = index_df[index_df['voice'] == 'ash']['parquet_file_name'].unique()[:5]
ds = load_dataset("humair025/Munch", data_files=list(target_files), split="train")
```

### Storage Optimization

```python
# If storage is limited, consider:
# 1. Download only specific voices
# 2. Download in batches and process incrementally
# 3. Use the hashed index for metadata-only analysis
# 4. Delete processed files after feature extraction
```

---

## 📜 Citation

If you use this dataset in your research, please cite:

### BibTeX

```bibtex
@dataset{munch_urdu_tts_2025,
  title={Munch: Large-Scale Urdu Text-to-Speech Dataset},
  author={Munir, Humair},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/humair025/Munch}},
  note={4.17M audio-text pairs across 13 voices}
}
```

### APA Format

```
Munir, H. (2025). Munch: Large-Scale Urdu Text-to-Speech Dataset [Dataset]. 
Hugging Face. https://huggingface.co/datasets/humair025/Munch
```

### MLA Format

```
Munir, Humair. "Munch: Large-Scale Urdu Text-to-Speech Dataset." Hugging Face, 2025, 
https://huggingface.co/datasets/humair025/Munch.
```

---

## 🤝 Contributing

Issues, suggestions, and contributions are welcome! Please:
- 🐛 Report data quality issues
- 💡 Suggest improvements
- 📝 Share your use cases and research
- 🔧 Contribute analysis scripts or tools

## 📄 License

This dataset is released under the **Creative Commons Attribution 4.0 International (CC-BY-4.0)** license.

You are free to:
- ✅ **Share** — copy and redistribute the material in any medium or format
- ✅ **Adapt** — remix, transform, and build upon the material for any purpose
- ✅ **Commercial use** — use the dataset for commercial purposes

Under the following terms:
- 📝 **Attribution** — You must give appropriate credit, provide a link to the license, and indicate if changes were made

---

## 🔗 Important Links

- 🎧 [**This Dataset (Full Audio)**](https://huggingface.co/datasets/humair025/Munch) - 1.27 TB
- 📊 [**Hashed Index**](https://huggingface.co/datasets/humair025/hashed_data) - ~1 GB metadata + hashes
- 🔄 [**Munch-1 (Newer Version)**](https://huggingface.co/datasets/humair025/munch-1) - 3.28 TB, 3.86M samples
- 💬 [**Discussions**](https://huggingface.co/datasets/humair025/Munch/discussions) - Ask questions, share research
- 🐛 [**Report Issues**](https://huggingface.co/datasets/humair025/Munch/discussions) - Data quality problems

---

## 🙏 Acknowledgments

- **TTS Generation**: OpenAI-compatible API endpoints
- **Voices**: 13 high-quality voice models (alloy, echo, fable, onyx, nova, shimmer, coral, verse, ballad, ash, sage, amuch, dan)
- **Infrastructure**: HuggingFace Datasets platform
- **Tools**: Python, datasets, pandas, numpy, scipy

---

## 📈 Usage Statistics

Help us understand how the dataset is used:
- Training TTS models
- Speech recognition research
- Voice cloning experiments
- Linguistic analysis
- Educational purposes
- Other (please share in discussions!)

---

## ⚡ Quick Start Tips

1. **First Time Users**: Start with the [hashed index](https://huggingface.co/datasets/humair025/hashed_data) (~1 GB) to explore the dataset
2. **Download Smart**: Use the index to find specific samples, then download only those parquet files
3. **Memory Matters**: Use streaming mode if working with large subsets
4. **Deduplication**: Check for duplicates using audio hashes before training
5. **Voice Selection**: Each voice has ~320k samples - choose based on your needs
6. **Consider Munch-1**: A newer version with 3.86M samples is also available at [humair025/munch-1](https://huggingface.co/datasets/humair025/munch-1)

---

**Note**: This is a large dataset (1.27 TB, 4.17M samples). Please download selectively based on your needs. Consider using the [hashed index](https://huggingface.co/datasets/humair025/hashed_data) for exploration and selective downloading.

**Last Updated**: December 2025

**Status**: ✅ Complete - All ~8,300 files uploaded

---

💡 **Pro Tip**: Download the lightweight [hashed index](https://huggingface.co/datasets/humair025/hashed_data) first to explore the dataset, find duplicates, and identify exactly which files you need - then download only those specific parquet files from this dataset!