Loading datasets...
Release Date: 7/20/2026
Format: TSV, MP3
Size: 3.83 GB
Share
This dataset is a derived, accent-filtered subset of the Mozilla Common Voice Scripted Speech 26.0 English release (`cv-corpus-26.0-2026-06-12`, locale `en`). It keeps only validated rows (present in the source `validated.tsv`) whose self-declared `accents` field is *exactly* `"India and South Asia (India, Pakistan, Sri Lanka)"`. This predefined Common Voice accent tag bundles India, Pakistan, and Sri Lanka. No gender or age filter is applied. The result is split speaker-disjoint into train/dev/test (no client_id appears in more than one split).
Restrictions/Special Constraints
N/A
Forbidden Usage
- You agree not to attempt to determine the identity of speakers. - You agree not to rehost this dataset.
Ethical Review
Source recordings and self-declared demographics (age, gender, accent) were volunteered by Common Voice contributors under Common Voice's own consent and contribution terms. This package only filters and repackages a subset; it does not collect new data or alter consent terms.
Intended Use
Research and development of English ASR systems, with emphasis on a South Asian accent slice.
Splits and speakers: 112,099 validated clips total — train 101,702 clips / 2,270 speakers, dev 6,533 clips / 102 speakers, test 3,864 clips / 81 speakers. Split is speaker-disjoint (no client_id appears in more than one split), proportioned 90.73/5.83/3.45 by clip count via a greedy per-speaker balancing algorithm targeting 95/2.5/2.5. The deviation from target is larger than in sibling subsets because per-speaker clip counts are extremely skewed here (max single speaker: 5,019 clips, ~4.5% of the whole subset; mean 45.7 clips/speaker, median 7), so the greedy assignment cannot track the target ratio as closely while keeping every speaker's clips within one split.
Clip duration: 166.00 hours total (597,605 seconds) across 112,099 clips. Per-split: train 149.95h, dev 10.12h, test 5.94h. Per-clip: mean 5.33s, median 5.23s, min 0.82s, max 27.79s, std 1.80s.
Gender distribution: male_masculine 67,235; female_feminine 38,917; blank/unspecified 5,947.
Age distribution (shipped subset, self-declared): twenties 73,613; thirties 12,323; fourties 10,458; blank/unspecified 5,854; teens 5,488; fifties 3,994; sixties 369.
Closed-form accent matching (QA): the accents field is pipe-delimited multi-select in the source corpus, but this subset was deliberately filtered to the closed-form (exact-match) rule — 100% of shipped clips have accents == exactly "India and South Asia (India, Pakistan, Sri Lanka)", no co-occurring tags.
Open-form vs. closed-form comparison (at execution time): open-form (accents contains "India and South Asia (India, Pakistan, Sri Lanka)") = 113,634 clips / 2,528 speakers / 168.46h; closed-form (accents == exactly that string) = 112,099 clips / 2,453 speakers / 166.00h. Closed-form drops 1,535 clips (1.35%).
Full-corpus accent x gender and age x gender cross-tabs are in stats.md.
Derivation methodology: derived from the Mozilla Common Voice Scripted Speech 26.0 English release (cv-corpus-26.0-2026-06-12, locale en). Filter: validated rows only (present in the source validated.tsv), whose self-declared accents field is exactly "India and South Asia (India, Pakistan, Sri Lanka)" (closed-form/exact match). No gender or age filter applied. Speaker-disjoint train/dev/test split via a greedy per-speaker assignment balancing toward 95/2.5/2.5 targets. Only the referenced MP3 clips were extracted from the ~2.58M-clip, ~89GB source archive (not the full archive) via a single sequential streaming pass over the source tar.gz.
