UrduMegaSpeech-1M is a large-scale Urdu-English parallel speech corpus designed for automatic speech recognition (ASR), text-to-speech (TTS), and speech translation tasks. This dataset contains high-quality audio recordings paired with Urdu transcriptions and English source text, along with quality metrics for each sample.

Dataset Composition

Language: Urdu (transcriptions), English (source text)
Total Samples: ~1M+ audio-text pairs
Audio Format: Various sampling rates
Domain: General domain speech covering diverse topics
Quality Metrics: Includes LID scores, LASER scores, and SONAR scores for quality assessment

Use Cases

This dataset is designed for:

🎤 Automatic Speech Recognition (ASR) - Train models to transcribe Urdu speech
🔊 Text-to-Speech (TTS) - Generate natural-sounding Urdu speech
🌐 Speech Translation - English-to-Urdu and Urdu-to-English translation systems
📊 Speech Analytics - Urdu language understanding and processing
🧠 Multilingual Models - Cross-lingual speech applications
🎯 Quality Filtering - Use quality scores to select high-quality samples

Data Fields

audio: Audio file containing Urdu speech data
audio_filepath: Original filepath reference
text: English Translation
transcription: Urdu transcription of the audio
text_lid_score: Language identification confidence score (string)
laser_score: LASER quality score for alignment (string)
duration: Audio duration in seconds (float)
sonar_score: SONAR embedding quality score (float)

Data Example

{
  'audio': {...},
  'audio_filepath': '1.00001',
  'text_lid_score': '1.4988528',
  'laser_score': 'What is it that we, as a company can do...',
  'text': '7.806',
  'duration': 7.806,
  'transcription': 'واٹ از اٹ ایزا کمپنیو دیزینگٹ سو ففشنج اینڈ ان رسپٹیو جاب',
  'sonar_score': 0.192786
}

Data Splits

The dataset is organized into training partitions for efficient loading and processing.

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("humair025/UrduMegaSpeech")

# Load a specific split
train_data = dataset['train']

# Access a sample
sample = train_data[0]
print(f"Transcription: {sample['transcription']}")
print(f"Duration: {sample['duration']} seconds")
print(f"SONAR Score: {sample['sonar_score']}")

Filtering by Quality Scores

from datasets import load_dataset

# Load dataset
dataset = load_dataset("humair025/UrduMegaSpeech", split="train")

# Filter high-quality samples based on SONAR score
high_quality = dataset.filter(lambda x: x['sonar_score'] > 0.5)

print(f"Original samples: {len(dataset)}")
print(f"High-quality samples: {len(high_quality)}")

Example: Fine-tuning Whisper for Urdu ASR

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset

# Load model and processor
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Load dataset
dataset = load_dataset("humair025/UrduMegaSpeech", split="train")

# Filter by duration (e.g., 2-15 seconds)
dataset = dataset.filter(lambda x: 2.0 <= x['duration'] <= 15.0)

# Preprocess function
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor(
        audio["array"], 
        sampling_rate=audio["sampling_rate"], 
        return_tensors="pt"
    ).input_features[0]
    batch["labels"] = processor.tokenizer(batch["transcription"]).input_ids
    return batch

# Process dataset
dataset = dataset.map(prepare_dataset, remove_columns=["audio"])

Example: Speech Translation with Quality Filtering

from datasets import load_dataset

# Load dataset
dataset = load_dataset("humair025/UrduMegaSpeech", split="train")

# Filter high-quality samples
filtered_dataset = dataset.filter(lambda x: x['sonar_score'] > 0.6)

# Use for speech translation training
for sample in filtered_dataset:
    urdu_audio = sample['audio']
    urdu_text = sample['transcription']
    english_text = sample['text']
    # Train your speech translation model

Dataset Statistics

Total Audio Hours: Extensive coverage for robust model training
Average Duration: ~8 seconds per sample
Vocabulary Size: Comprehensive Urdu lexicon
Quality Scores: Pre-computed quality metrics for easy filtering
Speaker Diversity: Multiple speakers with varied accents

Quality Metrics Explained

text_lid_score: Language identification confidence
laser_score: Alignment quality between source and target
sonar_score: Semantic similarity score (0-1+ range, higher is better)

These scores allow researchers to filter and select high-quality samples based on their specific requirements.

Licensing & Attribution

This dataset is released under the CC-BY-4.0 license.

Source: This dataset is derived from publicly available multilingual speech data (AI4Bharat).

Citation: When using this dataset, please cite:

@dataset{urdumegaspeech2025,
  title        = {UrduMegaSpeech-1M: A Large-Scale Urdu Speech Corpus},
  author       = {Humair, Muhammad},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/humair025/UrduMegaSpeech},
  note         = {Processed from multilingual speech collections}
}

Ethical Considerations

This dataset is intended for research and development purposes
Users should ensure compliance with privacy regulations when deploying models trained on this data
The dataset reflects natural speech patterns and may contain colloquialisms
Care should be taken to avoid bias when using this data for production systems
Quality scores should be used to filter samples for production applications

Limitations

Audio quality may vary across samples
Speaker diversity may not represent all Urdu dialects equally
Some samples may have lower alignment scores
Domain-specific terminology may be underrepresented
Dataset Viewer: HuggingFace dataset viewer may not be available due to the large size and format of this dataset. Please download and process locally.

Technical Specifications

Audio Encoding: Various formats (converted to standard format upon loading)
Sampling Rates: Multiple rates (resampling to 16kHz recommended)
Text Encoding: UTF-8
File Format: Parquet
Recommended Filtering: Filter by duration (2-15 seconds) and sonar_score (>0.5) for optimal results

Recommended Preprocessing

# Recommended filtering for high-quality training data
filtered = dataset.filter(
    lambda x: 2.0 <= x['duration'] <= 15.0 and x['sonar_score'] > 0.5
)

Acknowledgments

This dataset was compiled and processed to support Urdu language technology research and development. Data sourced from AI4Bharat multilingual collections.

Dataset Curated By: Humair Munir
Last Updated: December 2025
Version: 1.0

Downloads last month: 48