README.md

metadata

license: apache-2.0
task_categories:
  - text-to-speech
language:
  - ur
size_categories:
  - 10K<n<100K

Urdu TTS Dataset

A high-quality Urdu speech dataset for Text-to-Speech (TTS) model training, built from YouTube audiobooks and narrations.

Dataset Statistics

Property	Value
Language	Urdu (ur)
Total Samples	50000
Total Duration	70 hours
Sample Rate	22050 Hz
Audio Format	WAV (16-bit PCM, mono)

Columns

Column	Type	Description
audio	Audio	WAV file at 22050 Hz
transcript	string	Raw Urdu transcription (Whisper)
normalized_text	string	Cleaned and normalized Urdu text
audio_id	string	Unique segment identifier
duration_sec	float	Clip duration in seconds
snr_db	float	Estimated signal-to-noise ratio
source_file	string	Source YouTube video filename
language	string	Always "ur"

Usage

from datasets import load_dataset

dataset = load_dataset("{repo_id}")
print(dataset["train"][0])

Pipeline

1. YouTube audio download (yt-dlp)
2. Voice Activity Detection segmentation (1.5–15 sec clips)
3. SNR quality filtering
4. Urdu transcription (OpenAI Whisper medium/large)
5. Text normalization and deduplication
6. Incremental shard upload to HuggingFace

License

CC BY 4.0 — Please respect the rights of original content creators.