Common Voice Scripted Speech 26.0 - Scottish English

Splits and speakers: 20,293 validated clips total — train 19,211 clips / 188 speakers, dev 543 clips / 8 speakers, test 539 clips / 15 speakers. Split is speaker-disjoint (no client_id appears in more than one split), proportioned ~94.68/2.68/2.66 by clip count via a greedy per-speaker balancing algorithm targeting 95/2.5/2.5 (speakers have uneven clip counts, so exact ratios aren't achievable while keeping speakers disjoint).

Clip duration: 31.92 hours total (114,907 seconds) across 20,293 clips. Per-split: train 30.21h, dev 0.86h, test 0.85h. Per-clip: mean 5.66s, median 5.62s, min 1.10s, max 18.29s, std 1.98s.

Clip duration histogram:

   0-1s |                                                          0 ( 0.0%)
   1-2s | #####                                                  379 ( 1.9%)
   2-3s | #####################                                1,595 ( 7.9%)
   3-4s | ###################################                  2,624 (12.9%)
   4-5s | ###########################################          3,259 (16.1%)
   5-6s | ##################################################   3,739 (18.4%)
   6-7s | ############################################         3,311 (16.3%)
   7-8s | ###################################                  2,667 (13.1%)
   8-9s | ######################                               1,668 ( 8.2%)
  9-10s | ###########                                            840 ( 4.1%)
 10-12s | ##                                                     201 ( 1.0%)
 12-15s |                                                          8 ( 0.0%)
 15-20s |                                                          2 ( 0.0%)
 20-30s |                                                          0 ( 0.0%)
   30s+ |                                                          0 ( 0.0%)

Gender distribution: female_feminine 8,436; male_masculine 7,188; blank/unspecified 4,660; transgender 9.

Age distribution (shipped subset, self-declared): fourties 9,640; blank/unspecified 4,515; twenties 2,236; fifties 1,139; thirties 980; teens 745; sixties 603; seventies 435.

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 — 100.0000% of shipped clips have accents == exactly "Scottish English", no co-occurring tags. This is a deliberate divergence from the sibling British English submission, which used open-form (substring) matching and therefore carries a multi-tag distribution table; no such table is needed here.

Open-form vs. closed-form comparison (re-verified against live metadata at execution time): open-form (accents contains "Scottish English") = 20,481 clips / 239 speakers / 32.22h; closed-form (accents == exactly "Scottish English") = 20,293 clips / 211 speakers / 31.92h. Closed-form drops 188 clips (0.92%) — a small cost in exchange for removing ambiguity about whether a shipped clip's speaker also self-identified with a different accent.

Intentional overlap with British English submission: British English's region list (England, Scotland, Wales, (Northern) Ireland) intentionally includes Scotland, using open-form accent matching. A Scottish-tagged clip already shipped in british-english/packages/cv26-british-english.tar.gz will also ship here if it independently passes this subset's stricter closed-form filter. This overlap is intentional and confirmed with Oleg — the two submissions serve different accent-granularity use cases (broad British-Isles coverage vs. a clean, single-tag Scottish-only slice).

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 "Scottish English" (closed-form/exact match) — re-verified against live metadata at execution time to still hold (20,293 clips / 211 speakers / 31.92h, matching the pre-execution estimate). 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.