sajalmadan0909/hindi_transliteration

What’s in this repo

This dataset contains Hindi speech metadata + transliterations for multiple sources.
Audio is provided as tar shards (optimized upload) under data_archives/.

Main files (viewer-friendly)

• all_transliterated.jsonl: concatenation of the 3 JSONL files under transliterated/ (easy to browse in the Dataset Viewer)
• transliterated/*.jsonl: per-source JSONL with fields like:
  • dataset, id, index, audio_path
  • text / original_text
  • transliterated_text

Audio (optional)

To avoid uploading ~15k small .wav files, audio is stored as tar shards:

• data_archives/common_voice_hindi/audio_shard_000.tar ... audio_shard_006.tar
• data_archives/indicvoices_hindi/audio_shard_000.tar
• data_archives/hindi_fleurs/audio_shard_000.tar

Each tar contains files laid out like:

• <dataset_name>/audio/<id>.wav

How to download / use

Download metadata

Use the Hub UI or huggingface_hub to download all_transliterated.jsonl / transliterated/*.jsonl.

Download + extract audio shards

1. Download the shard(s) you need from data_archives/
2. Extract:

tar -xf data_archives/common_voice_hindi/audio_shard_000.tar

After extraction you’ll get a folder like common_voice_hindi/audio/*.wav which matches the relative audio_path structure used in the JSONL (with data/ replaced by your local extraction root as needed).

Notes

• This repo is private.
• The Dataset Viewer can take a few minutes to index after uploads. If it shows “available soon”, wait and refresh.