Tohirju/tajik-asr-ft-data

Tajik ASR — FT dataset (tg)

Compact Tajik ASR training set (16 kHz mono WAV), a smaller companion to the full set.

Train / Val	90,870 / 2,810 clips
Total audio	302.5 h, 16 kHz mono WAV
Frozen eval	fleurs_tg_test (600 clips — held out, NEVER train on it)
Audio	audio.tar — tar -xf audio.tar
Manifests	train_rel.jsonl / val_rel.jsonl / fleurs_tg_test_rel.jsonl (relative paths)

Preprocessing: transcript dedup, FLEURS 8-gram guard, duration 0.4–30 s.

⚠️ Revision note: the earliest revision (2026-08-01 ~13:49 UTC) carried incorrect manifests (a filter bug). Use the latest revision.

Public but gated (manual approval).

Licenses: constituent audio under CC-BY-4.0 and compatible terms; credit to the original corpus authors.