Urdu-LjSpeech Dataset

Dataset Description

Urdu-LjSpeech is a high-quality Urdu speech dataset designed for Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) tasks. The dataset contains Urdu audio recordings paired with their corresponding text transcriptions.

Dataset Summary

• Language: Urdu (اردو)
• Format: Audio files with text transcriptions
• Audio Specifications:
  • Sampling Rate: 22,050 Hz
  • Format: PCM 16-bit
  • Channels: Mono
• Use Cases: Text-to-Speech synthesis, Speech Recognition, Voice Cloning, Prosody Analysis

Supported Tasks

• Text-to-Speech (TTS): Train models to synthesize natural-sounding Urdu speech
• Automatic Speech Recognition (ASR): Develop speech-to-text systems for Urdu
• Voice Conversion: Train voice cloning and conversion models
• Linguistic Research: Study Urdu phonetics and prosody

Dataset Structure

Data Instances

Each instance in the dataset contains:

{
    'audio': {
        'array': array([...]), # Audio waveform
        'sampling_rate': 22050
    },
    'text': 'تم ہاتھ میں پتھر اٹھاتے ہو', # Urdu transcription
    'speaker': 'alloy', # Speaker identifier
    'id': 0 # Unique sample ID
}

Data Fields

Field	Type	Description
audio	Audio	Audio recording at 22,050 Hz sampling rate
text	string	Urdu text transcription in UTF-8
speaker	string	Speaker identifier
id	int	Unique identifier for the sample

Data Splits

The dataset is organized in batches for efficient loading:

dataset/
├── batch_0/
├── batch_1/
├── batch_2/
└── ...

Each batch contains approximately 1.5-2 GB of audio data.

Dataset Creation

Source Data

This dataset is a processed and validated version of speech recordings with careful quality control measures applied.

Data Collection

• Audio recordings were collected and validated for quality
• Each audio file was paired with its corresponding Urdu text transcription
• Quality validation includes:
  • Minimum audio duration check (>0.1 seconds)
  • PCM format validation
  • Corrupted audio removal
  • Text-audio alignment verification

Data Processing

The dataset underwent several processing steps:

1. Audio Validation: Each audio sample was validated for:
   • Sufficient duration (minimum 0.1 seconds)
   • Valid PCM format (even byte length for 16-bit samples)
   • No corruption or empty data
2. Batch Organization: Files organized into ~1.5-2GB batches for efficient streaming and downloading
3. Format Standardization: All audio normalized to:
   • 22,050 Hz sampling rate
   • 16-bit PCM format
   • Mono channel

Annotations

Text transcriptions are in standard Urdu script (UTF-8 encoded) with proper diacritical marks where applicable.

Usage

Loading the Dataset

from datasets import load_dataset
# Load the full dataset
dataset = load_dataset("humairawan/Urdu-LjSpeech")
# Load a specific batch
dataset = load_dataset("humairawan/Urdu-LjSpeech", data_dir="batch_0")
# Access samples
sample = dataset['train'][0]
print(sample['text']) # Print Urdu text
audio_array = sample['audio']['array'] # Access audio waveform
sampling_rate = sample['audio']['sampling_rate'] # Get sampling rate

Training a TTS Model

from datasets import load_dataset
from transformers import SpeechT5ForTextToSpeech, SpeechT5Processor
# Load dataset
dataset = load_dataset("humairawan/Urdu-LjSpeech")
# Initialize model and processor
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
# Your training code here...

Training an ASR Model

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
# Load dataset
dataset = load_dataset("humairawan/Urdu-LjSpeech")
# Initialize model and processor
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base")
# Your training code here...

Considerations

Ethical Considerations

• This dataset is intended for research and development of Urdu language technologies
• Users should be aware of potential biases in speaker representation
• Commercial use should respect speaker rights and consent

Citation

If you use this dataset in your research or applications, please cite it using the following BibTeX entry:

@dataset{awan2024urdu_ljspeech,
  author       = {Humair Munir},
  title        = {Urdu-LjSpeech: A High-Quality Urdu Speech Dataset for TTS and ASR},
  month        = dec,
  year         = 2024,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/humairawan/Urdu-LjSpeech}
}