mahwizzzz/Urdu-Streaming-ASR-Data

Urdu Streaming ASR — Combined Training Data Index

Reproducibility index for the curated Urdu ASR training mixture used in Urdu Streaming Zipformer 20M.

This release is a metadata-only index, not a downloadable audio or transcription corpus. To obtain the original recordings and transcripts, use the upstream owners' datasets under their own access and licensing terms.

Dataset at a glance

• 66,601 accepted unique training utterances; 131.0095 hours of original audio in the audited local training mixture.
• Four sources: human-labeled read, conversational and expressive speech plus screened silver/code-switched labels.
• Training rows only. No sealed dev/test audio, references, or benchmark transcripts are republished.
• Counts are unique accepted training records; a balanced training sampler can repeat records without creating new audio.

Sources and provenance

Source	Unique records	Accepted hours	Source licensing and access notes
IndicVoices Urdu	34,920	69.7864	CC-BY-4.0; gated access
KathBath Urdu	12,267	18.0598	Hub CC-BY-4.0; source-text rights caveat
Rasa Urdu	1,023	1.8653	CC-BY-4.0; gated access
UrduSpeech	18,391	41.2980	CC-BY-4.0 declared; underlying media rights require review

Source revisions, upstream links and counts are also provided in sources.json. See upstream terms and redistribution policy.

Included files

• data/train.jsonl — one non-audio, non-transcript provenance row per accepted training utterance.
• SUMMARY.json — audited counts and source breakdown.
• sources.json — upstream source and revision inventory.
• SHA256SUMS — local integrity hashes for the staged data and metadata.

Manifest fields: sample_id (one-way derived digest of the canonical local ID), source, source_revision, split, duration_seconds, label_origin, domain, language_mix, license_id_from_source. It deliberately omits filenames, personal identifiers, recording references, local storage paths and all transcription content.

Reproducibility

GitHub — code, training recipes, configs, data-source inventory and evaluation

• Data sources
• Data quality and leakage methodology
• Source configuration with pinned revisions
• Model weights and inference

The index is not sufficient on its own to reconstruct copyrighted/gated recordings, and does not replace obtaining authorized upstream data. The verified model's independent evaluation partitions remain excluded.

Example: inspect metadata

import json
with open('data/train.jsonl', encoding='utf-8') as file:
    for line in file:
        row = json.loads(line)
        print(row['source'], row['duration_seconds'], row['label_origin'])
        break

Ethics and limitations

The training mixture contains conversational speech, expressive speech, and supervised silver labels. Original sources have different privacy, licensing and access conditions. Do not attempt speaker identification or re-identification. Silver labels are not equivalent to human-verified gold. No claim is made that every original audio clip can be re-hosted under a single license. This release does not contain audio files or transcriptions.

Citation and attribution

Cite each upstream dataset used as well as this derivative curation index, with the references documented by the corresponding source owner. The source code's Apache-2.0 license does not relicense the upstream audio.