theusamaaslam/urdu-organic-collection

Urdu Organic Speech Collection

Consolidated organic Urdu speech dataset containing 337,874 clips (~337.2 hours) in one unified repository: training (326,923 clips / ~323.5h), validation (5,609 clips / ~6.6h), and a locked test split (5,342 clips / ~7.1h).

The corpus combines the previously published organic Urdu collection with the audited 208h organic Urdu corpus — a large, translator-independent organic dataset that was speaker-labeled and quality-audited before inclusion. Source provenance and speaker identity remain available as row-level columns.

Quick facts

Split	Clips	Hours
train	326,923	~323.5h
validation	5,609	~6.6h
test	5,342	~7.1h
Total	337,874	~337.2h

The training split is published as one consolidated body; full per-source counts are in dataset_stats.json.

Included sources

• ghalib_208h_organic: the audited 208h organic Urdu corpus (259,270 train clips, ~245.83h, 442 speakers). Each clip was speaker-clustered (CAMPPlus embeddings), signal-quality audited (silence, quietness, clipping, noise, edge truncation), and transcript-alignment audited (median CER ≈ 8% on a 3.5k-clip sample).
• collective: cleaned Common Voice Urdu plus the two previously incorporated Hugging Face Common Voice derivatives. Speaker-disjoint validation and locked test splits are preserved.
• permitted_10h: the authorized local multispeaker corpus after one-to-one transcript matching and filtering.
• fleurs: Google FLEURS ur_pk, preserving train, validation, and test splits.
• uat: a pinned copy of mahwizzzz/UAT at revision da45a8fce76781c4b8e9609f256d737d5dc5f91d.

All sources are merged into train, validation, and test. The source column enables duration-based sampling; speaker_id enables speaker-aware batching or speaker-disjoint evaluation.

The partially downloaded ASLP pilot is intentionally excluded: no final, audited training manifest was produced, and publishing it as training-ready data would misrepresent its status.

Schema

id, audio, text, speaker_id, source, duration_seconds, language, is_synthetic. Audio is embedded in Parquet with the text always present in the shard; id is unique per row.

Readable transcriptions are also provided under metadata/:

• metadata/train.jsonl — original collective train rows
• metadata/train_208h.jsonl — the added 208h organic rows
• metadata/uat_train.jsonl — UAT rows
• metadata/validation.jsonl, metadata/test.jsonl — held-out splits

Usage

from datasets import load_dataset

organic = load_dataset("theusamaaslam/urdu-organic-collection", token=True)
print(organic)

Source terms and licensing

This is a private, mixed-license collection. Repository access does not replace the upstream source terms. Common Voice-derived data remains subject to its upstream terms; FLEURS is published upstream under CC-BY-4.0; the permitted local corpus is restricted to the authorization supplied by its owner; the UAT dataset card declares no license, so its redistribution and production use require separate provenance review.

Training guidance

Use duration-based source sampling rather than treating all sources as equal-quality, keep validation/test partitions locked, deduplicate against any target voice corpus, and audit transcript alignment before each training run.