The Urdu-aud0 dataset is a high-quality collection of synthetic Urdu audio samples paired with transcripts, designed for Text-to-Speech (TTS) model training, evaluation, and fine-tuning. The dataset contains approximately 45,000 audio clips generated using OpenAI's Audio API with the "dan" voice model.

Each entry includes:

High-quality audio in WAV format (22,050 Hz, mono, 16-bit)
Corresponding Urdu text transcripts
Generation timestamps
Voice model identifier

With approximately 147 hours (98 × 1.5 hours) of total audio duration, this substantial dataset addresses the scarcity of Urdu speech data, making it valuable for developing speech technologies in this low-resource language. The audio has been validated for quality, with corrupted entries removed and proper WAV headers added during processing.

Supported Tasks

Primary Tasks

Text-to-Speech (TTS): Train or fine-tune TTS models to generate natural-sounding Urdu speech
Automatic Speech Recognition (ASR): Train or evaluate ASR models on Urdu audio-transcript pairs
Speech Synthesis Evaluation: Benchmark using metrics like Mean Opinion Score (MOS) or Word Error Rate (WER)

Secondary Tasks

Language Modeling: Urdu text transcripts can support language model training
Prosody Analysis: Study speech patterns in synthetic Urdu audio
Voice Cloning Research: Reference data for Urdu voice synthesis
Cross-voice Model Comparison: Compare with other voice models (e.g., "sage" in Urdu-aud01)

Languages

Primary Language: Urdu (ur)
Script: Urdu script (Nastaliq/Arabic-based)
Variety: Modern Standard Urdu
Audio Source: OpenAI Audio API (voice: "dan")

All transcripts are in Urdu script, and audio follows standard Urdu pronunciation patterns.

Dataset Structure

Data Instances

Each instance represents a single audio clip with metadata:

{
  "id": 1,
  "audio": {
    "path": null,
    "array": null,
    "sampling_rate": 22050
  },
  "transcript": "\"یہ ایک نمونہ جملہ ہے۔\"",
  "voice": "dan",
  "text": "یہ ایک نمونہ جملہ ہے۔",
  "timestamp": "2025-12-01T10:30:45.123456",
  "error": null
}

Data Fields

Field	Type	Description
`id`	`int64`	Unique identifier for the audio sample
`audio`	`audio`	Audio data dictionary with bytes (WAV format) at 22,050 Hz sampling rate
`transcript`	`string`	Quoted version of the transcript text
`voice`	`string`	Voice model used (OpenAI "dan" voice)
`text`	`string`	Clean, unquoted Urdu transcript
`timestamp`	`string`	ISO-formatted generation timestamp
`error`	`string`	Error message if any (mostly null)

Audio Specifications

Format: WAV (RIFF header)
Sampling Rate: 22,050 Hz
Channels: Mono (1)
Bit Depth: 16-bit PCM
Minimum Duration: ~0.1 seconds per clip
Total Duration: Approximately 147 hours (98 × 1.5 hours)
Average Duration: ~11.76 seconds per clip (147 hours ÷ 45,000 clips)

Data Splits

Train: ~45,000 samples (full dataset)
No predefined validation or test splits

Recommendation: Create custom splits (e.g., 80/10/10 for train/validation/test) based on your specific requirements.

Dataset Creation

Curation Rationale

This dataset was created to address the critical shortage of high-quality Urdu speech data for TTS and ASR applications. By leveraging OpenAI's Audio API with the "dan" voice model, we generated consistent, high-quality synthetic speech that can serve as training data for various speech processing tasks.

Key objectives:

Provide reliable Urdu audio-text pairs for model training
Ensure data quality through rigorous validation
Support development of Urdu language technologies
Enable research in low-resource language speech processing
Offer voice diversity (complement to "sage" voice in Urdu-aud01)

Source Data

Generation Method: Synthetic audio generated using OpenAI's Audio API (voice model: "dan")

Processing Pipeline:

Source data loaded from humair025/aud0 repository
Audio validation (minimum length, byte count verification)
WAV header addition to raw PCM data
Corrupted entries removed
Batching by size (1.5-2GB per batch)
Upload to Hugging Face Hub

Processing Script: Custom Python pipeline using:

Hugging Face Datasets library
Pandas for data manipulation
Hugging Face Hub API for uploads

Annotations

Annotation Type: Automated during TTS generation

Annotators: OpenAI Audio API system (no human annotation)

Quality Assurance:

Automated validation removes audio files smaller than 0.1 seconds
Even byte count verification for PCM data
WAV header verification after conversion
Integrity checks during batching process

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("humairawan/Urdu-aud0")

# Access a sample
sample = dataset["train"][0]
print(f"Text: {sample['text']}")
print(f"Voice: {sample['voice']}")
print(f"Audio sampling rate: {sample['audio']['sampling_rate']}")

Creating Train/Val/Test Splits

from datasets import load_dataset

dataset = load_dataset("humairawan/Urdu-aud0")

# Create 80/10/10 splits
train_test = dataset["train"].train_test_split(test_size=0.2, seed=42)
test_val = train_test["test"].train_test_split(test_size=0.5, seed=42)

splits = {
    "train": train_test["train"],      # ~36,000 samples
    "validation": test_val["train"],    # ~4,500 samples
    "test": test_val["test"]           # ~4,500 samples
}

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

Example: Fine-tuning a TTS Model

from datasets import load_dataset

# Load dataset
dataset = load_dataset("humairawan/Urdu-aud0", split="train")

# Prepare for training
def prepare_example(example):
    return {
        "text": example["text"],
        "audio": example["audio"]["array"],
        "sampling_rate": example["audio"]["sampling_rate"]
    }

prepared_dataset = dataset.map(prepare_example)

Combining Multiple Voice Models

from datasets import load_dataset, concatenate_datasets

# Load datasets with different voices
aud0_dan = load_dataset("humairawan/Urdu-aud0", split="train")     # dan voice
aud01_sage = load_dataset("humairawan/Urdu-aud01", split="train")  # sage voice

# Combine for multi-voice training
combined = concatenate_datasets([aud0_dan, aud01_sage])
print(f"Combined dataset: {len(combined)} samples")

Considerations for Using the Data

Social Impact

Positive Impacts:

Enhances accessibility for Urdu speakers through voice assistants and screen readers
Supports AI development in underrepresented languages
Enables educational tools for Urdu language learning
Facilitates content creation in Urdu
Provides substantial hours of training data (147 hours)

Potential Concerns:

Synthetic data may not capture full range of human speech variation
Limited to single voice model "dan" (may not represent dialectal diversity)
Should be supplemented with natural speech data for production systems

Bias Considerations

Voice Diversity: All audio generated using a single voice model ("dan"), which limits:

Accent representation
Gender diversity (male-leaning voice characteristics)
Age range variation
Emotional expression variety

Content Bias: Text content reflects the source material and may have:

Topic skew towards certain domains
Formality bias in language use
Potential cultural or regional biases

Recommendations:

Combine with natural speech datasets when possible
Use alongside Urdu-aud01 (sage voice) for voice diversity
Be aware of synthetic nature when deploying in real-world applications
Test models on diverse natural speech samples before production deployment

Known Limitations

Synthetic Nature: Does not capture real-world acoustic conditions, background noise, or natural prosody variations
Single Voice: Limited diversity in voice characteristics (all "dan" voice)
Language Variety: Focuses on Modern Standard Urdu; may not represent all dialects
No Speaker Metadata: No demographic information about hypothetical speakers
Error Field: Some entries may have error flags (check before use)
Voice Gender: "Dan" voice has male-leaning characteristics, limiting gender representation

Ethical Considerations

This dataset uses synthetic audio generated by OpenAI's API, which is subject to OpenAI's terms of service
Users should be transparent about using synthetic data when deploying models
For applications requiring natural speech, supplement with human-recorded data
Consider privacy implications when generating similar synthetic datasets
Be mindful of potential bias towards male voice characteristics

Additional Information

Dataset Curators

Primary Curator: Humair Munir (@humairawan)
Original Source: humair025 (@humair025)

Licensing Information

License: Apache 2.0

This dataset is released under the Apache License 2.0, which permits:

Commercial use
Modification
Distribution
Private use

See the LICENSE file for full terms.

Note: Audio generated using OpenAI's Audio API is subject to OpenAI's terms of service and usage policies.

Citation Information

If you use this dataset in your research or projects, please cite it as:

@dataset{humairmunir_urdu_aud0_2025,
  author = {Humair Munir},
  title = {Urdu-aud0: Synthetic Urdu TTS Audio Dataset with Dan Voice},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/humairawan/Urdu-aud0}},
  note = {Audio generated using OpenAI Audio API (dan voice), 147 hours}
}

Acknowledgments

OpenAI for providing the Audio API used to generate this dataset
The Hugging Face team for hosting infrastructure and tools
The original data curators and processors
The Urdu language community for supporting low-resource language development

Contributions

We welcome contributions to improve this dataset:

Issue Reports: Report problems or suggest improvements via GitHub Issues
Pull Requests: Submit improvements to processing scripts or documentation
Extensions: Consider creating complementary datasets (e.g., different voice models, domains)

Contact: For questions or collaboration, reach out through Hugging Face discussions or create an issue in the repository.

Version History

v1.0 (December 2025): Initial release with ~45,000 validated audio samples
- Audio source: OpenAI Audio API (dan voice)
- Format: WAV, 22,050 Hz, mono, 16-bit
- Total duration: ~147 hours
- Processing: Validation, WAV header addition, batching

Related Resources

Urdu-aud01 Dataset: Companion dataset with "sage" voice (~59,000 samples)

Dataset Statistics

Metric	Value
Total Samples	~45,000 +
Voice Model	dan (OpenAI)
Total Duration	~147 hours
Average Clip Length	~11.76 seconds
Sampling Rate	22,050 Hz
Audio Format	WAV (16-bit PCM mono)
Language	Urdu (ur)
License	Apache 2.0

Last Updated: December 2025

Downloads last month: 17