khawajaaliarshad/common-voice-urdu-processed-expanded

🎙️ Common Voice Urdu (Processed & Expanded)

The largest ready-to-use Urdu speech dataset for fine-tuning ASR models

Mozilla Common Voice → Preprocessed & Expanded → Whisper-Ready ✨

---

📊 Dataset at a Glance

Split	Samples	Use
🏋️ Train	54,891	Model training
🔧 Validation	5,000	Hyperparameter tuning
🧪 Test	5,000	Final evaluation
Total	64,891

💡 Audio is pre-resampled to 16kHz — plug directly into Whisper!

📈 3.7x more training data than the original processed dataset!

---

🚀 Quick Start

from datasets import load_dataset

One line to load!

ds = load_dataset("khawajaaliarshad/common-voice-urdu-processed-expanded")

Peek at the data

sample = ds["train"][0] print(f"🗣️ {sample['sentence']}")

Audio ready at: sample["audio"]["array"]

🔥 Whisper Fine-tuning Example

from transformers import WhisperProcessor

processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="ur")

def prepare(batch): batch["input_features"] = processor.feature_extractor( batch["audio"]["array"], sampling_rate=16000 ).input_features[0] batch["labels"] = processor.tokenizer(batch["sentence"]).input_ids return batch

ds = ds.map(prepare, remove_columns=ds["train"].column_names)

---

📋 Data Fields

Field	Type	Description
audio	🔊 Audio	16kHz waveform
sentence	📝 String	Urdu transcription
up_votes / down_votes	👍👎 Int	Quality votes
age, gender, accent	👤 String	Speaker metadata

---

🙏 Credits

🎙️ Mozilla Common Voice	📦 Mozilla Data Collective
Open speech dataset	API access

Thanks to all Urdu-speaking contributors who donated their voices! 🎤

---

📄 License: CC0-1.0 (Public Domain) — Use freely for anything!

Made with ❤️ for اردو