Datasets:

humair025
/

munch_urdu_preview

license: cc-by-4.0
task_categories:
  - text-to-speech
  - automatic-speech-recognition
language:
  - ur
tags:
  - Urdu
  - TTS
  - ASR
  - speech-synthesis
  - preview
  - audio
size_categories:
  - 10K<n<100K
pretty_name: Munch Preview

🎧 Munch Preview Dataset

📖 Table of Contents

Dataset Description
Dataset Structure
Dataset Creation
Usage
Considerations
Citation
Contact

📋 Dataset Description

Overview

Munch Preview is a carefully curated preview dataset containing ** high-quality Urdu text-to-speech samples** from both versions of the Munch dataset family. This lightweight version allows researchers, developers, and practitioners to quickly explore and prototype with Urdu TTS data without downloading the full multi-terabyte datasets.

Purpose

This preview dataset serves multiple purposes:

🎯 Quick Exploration: Rapidly assess the quality and characteristics of Munch datasets
🧪 Prototyping: Test TTS/ASR pipelines before committing to full dataset downloads
📚 Educational: Learn about Urdu speech synthesis with manageable data sizes
🔬 Algorithm Development: Develop and validate algorithms on representative samples
📊 Comparative Analysis: Compare v1 and v2 dataset characteristics side-by-side

Key Features

✅ Two Dataset Versions: Samples from both Munch v1 and Munch-1 v2
✅ Balanced Sampling: Stratified random sampling across all 13 voices
✅ WAV Format: High-quality audio playable directly in HuggingFace viewer
✅ Fast Download: ~4.36 GB vs 4+ TB for full datasets
✅ Production Quality: Same preprocessing and quality as full datasets
✅ Metadata Rich: Complete transcripts, timestamps, and error tracking

Languages

Primary: Urdu (ur)
Script: Arabic script (Nastaliq)

📊 Dataset Structure

Data Instances

Each sample in the dataset contains:

{
    'id': 123456,
    'text': 'یہ ایک نمونہ متن ہے',
    'transcript': 'یہ ایک نمونہ متن ہے',
    'voice': 'ash',
    'audio': {
        'array': array([...]),  # Audio waveform
        'path': None,
        'sampling_rate': 22050
    },
    'timestamp': '2025-12-01T10:30:45',
    'error': None,
    'original_parquet': 'train-00123.parquet',
    'dataset_version': 'munch_v1'
}

Data Splits

Split	Samples	Size	Source	Description
munch_v1	~4.9k	~1.68 GB	Munch v1	Samples from original Munch dataset (4.17M total)
munch_v2	~4k	~2.68 GB	Munch-1 v2	Samples from improved Munch-1 dataset (3.86M total)
Total	~8k	~4.36 GB	-	Combined preview dataset

Data Fields

Field	Type	Description
`id`	`int`	Unique paragraph identifier from source dataset
`text`	`string`	Original Urdu text (input to TTS system)
`transcript`	`string`	Transcription of generated audio (may differ from input)
`voice`	`string`	Voice identifier (13 options per dataset)
`audio`	`Audio`	Audio data with waveform array and metadata
`timestamp`	`string`	ISO 8601 timestamp of audio generation
`error`	`string`	Error message if generation failed (usually None)
`original_parquet`	`string`	Source parquet file from full dataset
`dataset_version`	`string`	Version identifier: "munch_v1" or "munch_v2"

Audio Specifications

Format: WAV (Waveform Audio File Format)
Sample Rate: 22,050 Hz
Channels: Mono (1 channel)
Bit Depth: 16-bit signed integer PCM
Encoding: Linear PCM
Average Duration: X seconds per sample
Total Duration: ~X hours (combined)

Voice Distribution

Each split contains samples from 13 different voices with approximately balanced distribution (~770 samples per voice):

Available Voices:

alloy - Neutral, clear voice
echo - Resonant, deep voice
fable - Storytelling voice
onyx - Strong, authoritative voice
nova - Bright, energetic voice
shimmer - Soft, gentle voice
coral - Warm, friendly voice
verse - Poetic, expressive voice
ballad - Melodic, smooth voice
ash - Natural, conversational voice
sage - Wise, measured voice
amuch - Custom voice variant
dan - Custom voice variant

🔨 Dataset Creation

Source Data

Initial Data Collection

This preview dataset is derived from two large-scale Urdu TTS datasets:

Munch v1 (humair025/Munch)
- Total Size: 1.27 TB
- Total Samples: 4,167,500
- Collection Period: 2025
Munch-1 v2 (humair025/munch-1)
- Total Size: 3.28 TB
- Total Samples: 3,856,500
- Collection Period: 2025-2025

Data Pipeline

The source datasets were created using:

Urdu text corpus from various domains (literature, news, social media, technical)
State-of-the-art neural TTS synthesis
Multiple voice profiles for diversity
Quality validation and error tracking

Annotations

Annotation Process

Automatic: All audio was generated using text-to-speech systems
Transcription: Generated transcripts may differ slightly from input text due to TTS normalization
Error Tracking: Samples with generation errors are flagged in the error field
No Human Annotation: This is a synthetic dataset with automatic metadata

Personal and Sensitive Information

No Personal Information: All text and audio are synthetically generated
No Speaker Identification: Voices are synthetic and do not correspond to real individuals
No Biometric Data: Audio is generated, not recorded from human speakers

💻 Usage

Loading the Dataset

Basic Loading

from datasets import load_dataset

# Load complete dataset (both splits)
dataset = load_dataset("humair025/munch_preview")

print(f"Munch v1: {len(dataset['munch_v1']):,} samples")
print(f"Munch v2: {len(dataset['munch_v2']):,} samples")

Load Specific Split

# Load only v1
v1_dataset = load_dataset("humair025/munch_preview", split="munch_v1")

# Load only v2
v2_dataset = load_dataset("humair025/munch_preview", split="munch_v2")

Streaming Mode

# For even lower memory usage
dataset = load_dataset("humair025/munch_preview", streaming=True)

for sample in dataset['munch_v1']:
    print(sample['text'])
    break

Basic Usage Examples

1. Audio Playback (Jupyter/Colab)

import IPython.display as ipd

# Play first sample from v1
sample = dataset['munch_v1'][0]
print(f"Text: {sample['text']}")
print(f"Voice: {sample['voice']}")

ipd.display(ipd.Audio(
    sample['audio']['array'], 
    rate=sample['audio']['sampling_rate']
))

2. Export to WAV Files

import soundfile as sf
import os

# Export first 10 samples
os.makedirs("audio_samples", exist_ok=True)

for i, sample in enumerate(dataset['munch_v1'][:10]):
    filename = f"audio_samples/v1_sample_{i:03d}_{sample['voice']}.wav"
    sf.write(
        filename, 
        sample['audio']['array'], 
        sample['audio']['sampling_rate']
    )
    print(f"Saved: {filename}")

3. Filter by Voice

# Get all samples from specific voice
ash_samples = [
    sample for sample in dataset['munch_v1'] 
    if sample['voice'] == 'ash'
]

print(f"Found {len(ash_samples)} samples with 'ash' voice")

4. Analyze Text Statistics

import pandas as pd

# Convert to DataFrame for analysis
df_v1 = pd.DataFrame(dataset['munch_v1'])
df_v2 = pd.DataFrame(dataset['munch_v2'])

print("Text Length Statistics (v1):")
print(df_v1['text'].str.len().describe())

print("\nText Length Statistics (v2):")
print(df_v2['text'].str.len().describe())

5. Audio Duration Analysis

import numpy as np

# Calculate durations
v1_durations = [
    len(sample['audio']['array']) / sample['audio']['sampling_rate']
    for sample in dataset['munch_v1']
]

v2_durations = [
    len(sample['audio']['array']) / sample['audio']['sampling_rate']
    for sample in dataset['munch_v2']
]

print(f"V1 average duration: {np.mean(v1_durations):.2f}s")
print(f"V2 average duration: {np.mean(v2_durations):.2f}s")
print(f"V1 total duration: {sum(v1_durations)/3600:.2f} hours")
print(f"V2 total duration: {sum(v2_durations)/3600:.2f} hours")

Advanced Usage

1. Compare Dataset Versions

from collections import Counter

# Voice distribution comparison
v1_voices = Counter([s['voice'] for s in dataset['munch_v1']])
v2_voices = Counter([s['voice'] for s in dataset['munch_v2']])

print("Voice Distribution Comparison:")
print(f"{'Voice':<10} {'V1 Count':<10} {'V2 Count':<10}")
print("-" * 30)
for voice in sorted(v1_voices.keys()):
    print(f"{voice:<10} {v1_voices[voice]:<10} {v2_voices[voice]:<10}")

2. Train/Validation Split

from datasets import DatasetDict

# Split v1 into train/validation (80/20)
v1_split = dataset['munch_v1'].train_test_split(test_size=0.2, seed=42)

train_val_dataset = DatasetDict({
    'train': v1_split['train'],
    'validation': v1_split['test']
})

print(f"Train: {len(train_val_dataset['train'])} samples")
print(f"Validation: {len(train_val_dataset['validation'])} samples")

3. Combine Both Versions

from datasets import concatenate_datasets

# Combine v1 and v2 for training
combined = concatenate_datasets([
    dataset['munch_v1'], 
    dataset['munch_v2']
])

print(f"Combined dataset: {len(combined):,} samples")

# Shuffle for training
combined_shuffled = combined.shuffle(seed=42)

4. Feature Extraction

import librosa

# Extract MFCC features from audio
def extract_features(sample):
    audio_array = sample['audio']['array']
    sr = sample['audio']['sampling_rate']
    
    # Extract MFCCs
    mfccs = librosa.feature.mfcc(
        y=audio_array.astype(float), 
        sr=sr, 
        n_mfcc=13
    )
    
    return {
        'mfcc_mean': mfccs.mean(axis=1),
        'mfcc_std': mfccs.std(axis=1)
    }

# Apply to first 100 samples
features = [extract_features(s) for s in dataset['munch_v1'][:100]]

5. Create Evaluation Set

# Create balanced evaluation set (50 samples per voice)
eval_samples = []

for voice in ['alloy', 'echo', 'fable', 'onyx', 'nova']:
    voice_samples = [
        s for s in dataset['munch_v1'] 
        if s['voice'] == voice
    ][:50]
    eval_samples.extend(voice_samples)

print(f"Evaluation set: {len(eval_samples)} samples")

🔍 Considerations for Using the Data

Social Impact

Positive Impacts

Language Preservation: Supports Urdu language technology development
Accessibility: Enables text-to-speech applications for Urdu speakers
Research Enablement: Provides researchers with quality Urdu audio data
Educational: Facilitates Urdu language learning applications
Low Barrier: Small size enables experimentation without major compute resources

Potential Concerns

Synthetic Bias: Audio is synthetic and may not represent natural Urdu speech patterns
Voice Diversity: Limited to 13 voice profiles, may not represent full spectrum of Urdu speakers
Domain Coverage: Text sources may not cover all Urdu dialects or specialized domains
Quality Variance: As a preview, sampling may not perfectly represent full dataset quality distribution

Discussion of Biases

Known Biases

Synthetic Speech Bias: All audio is TTS-generated, not natural speech
- May contain artifacts specific to TTS systems
- Prosody and intonation may differ from human speech
Voice Selection: 13 voices may not represent:
- Full range of Urdu accents (Pakistani, Indian variations)
- Age diversity (child, elderly speakers)
- Regional dialects
Text Domain: Source text may be biased toward:
- Formal/written Urdu vs. colloquial speech
- Certain topics or domains
- Modern vocabulary vs. classical Urdu
Sampling Bias: Preview sampling may:
- Over/under-represent certain characteristics
- Not capture edge cases present in full dataset

Limitations

Technical Limitations

Preview Size: Only ~0.25% of full datasets
Voice Coverage: 13 voices may be insufficient for some applications
Quality Variance: Random sampling may include quality outliers

Use Case Limitations

Suitable For:

✅ Quick prototyping and testing
✅ Algorithm development
✅ Educational purposes
✅ Pipeline validation
✅ Quality assessment

Not Suitable For:

❌ Production model training (use full datasets)
❌ Comprehensive benchmarking
❌ Statistical significance testing
❌ Fine-grained quality analysis

Recommendations

For Research: Use full datasets (Munch v1 or v2) for final experiments
For Production: Validate on full dataset before deployment
For Training: Consider this as dev/test set, use full dataset for training
For Evaluation: Supplement with natural Urdu speech data

Privacy and Ethics

No Privacy Concerns: Fully synthetic data with no personal information
No Consent Required: No human speakers involved
Ethical Considerations:
- Synthetic voices should be clearly labeled as such in applications
- Consider potential misuse for deepfakes or impersonation
- Respect Urdu language and culture in applications

📈 Dataset Statistics

Overall Statistics

Metric	Value
Total Samples	~20,000
Total Size	~1.2 GB
Audio Duration	~16-28 hours
Languages	1 (Urdu)
Voices	13 per split
Sample Rate	22,050 Hz
Bit Depth	16-bit
Average Sample Duration	3-5 seconds

Per-Split Statistics

Metric	Munch v1	Munch v2
Samples	~10,000	~10,000
Size	~600 MB	~600 MB
Source Dataset Size	1.27 TB	3.28 TB
Source Total Samples	4,167,500	3,856,500
Sampling Rate	~0.24%	~0.26%

Text Statistics (Estimated)

Metric	Range
Characters per Sample	20-200
Words per Sample	5-40
Average Text Length	~50-80 characters

🔗 Related Datasets

Full Datasets

Dataset	Size	Samples	Link
Munch v1	1.27 TB	4.17M	humair025/Munch
Munch-1 v2	3.28 TB	3.86M	humair025/munch-1

Index Datasets (Metadata Only)

Dataset	Size	Purpose	Link
Munch v1 Index	~1 GB	Fast exploration without audio	humair025/hashed_data
Munch v2 Index	~1 GB	Fast exploration without audio	humair025/hashed_data_munch_1

Upgrade Path

Preview Dataset (4.36 GB)
    ↓
Test on preview, validate approach
    ↓
Index Datasets (~1 GB each)
    ↓
Explore metadata, plan subsets
    ↓
Full Datasets (1.27 TB / 3.28 TB)
    ↓
Production training

📜 Licensing Information

License

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

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, even commercially

Under the following terms:

Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made

Full License Text

Citation Requirements

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

@dataset{munch_preview_2025,
  title={Munch Preview: Quick Start Urdu Text-to-Speech Dataset},
  author={ Humair Munir},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/humair025/munch_preview}},
  note={Preview subsets of Munch v1 and Munch-1 v2 datasets}
}

For the full datasets, also cite:

@dataset{munch_v1_2025,
  title={Munch: Large-Scale Urdu Text-to-Speech Dataset},
  author={ Humair Munir},
  author={ humair025},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/humair025/Munch}}
}

@dataset{munch_v2_2025,
  title={Munch-1: Large-Scale Urdu Text-to-Speech Dataset},
  author={ Humair Munir},
  author={ humair025}
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/humair025/munch-1}}
}

👥 Dataset Curators

Created by: Humair Munir

Organization: Independent

Contact: Available through HuggingFace dataset page

🙏 Acknowledgements

This preview dataset is made possible by:

The creation of the full Munch and Munch-1 datasets
HuggingFace for dataset hosting infrastructure
Open-source TTS technology developers

📞 Contact & Support

Questions or Issues?

Dataset Issues: Use the Discussions tab
Feature Requests: Open an issue in Discussions
Bug Reports: Report in Discussions with detailed information

Additional Resources

Documentation: This README
Full Datasets: See Related Datasets section above
Community: Join discussions on HuggingFace

📅 Changelog

Version 1.0 (December 2025)

Initial release
~4k samples from Munch v1
~4k samples from Munch-1 v2
WAV format audio at 22,050 Hz
Complete metadata and documentation

⚡ Quick Reference

At a Glance

# Installation
pip install datasets soundfile

# Load
from datasets import load_dataset
ds = load_dataset("humair025/munch_preview")

# Explore
print(ds)
print(ds['munch_v1'][0])

# Play audio (Jupyter)
import IPython.display as ipd
sample = ds['munch_v1'][0]
ipd.display(ipd.Audio(sample['audio']['array'], rate=22050))

Key URLs

This Dataset: https://huggingface.co/datasets/humair025/munch_preview
Munch v1 Full: https://huggingface.co/datasets/humair025/Munch
Munch v2 Full: https://huggingface.co/datasets/humair025/munch-1
License: https://creativecommons.org/licenses/by/4.0/

Last Updated: December 2025