You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Tajik ASR — FULL dataset (tg)

Large training set for Tajik automatic speech recognition, 16 kHz mono WAV.

Train / Val 231,175 / 7,149 clips
Total audio 1,134.7 h, 16 kHz mono WAV
Frozen eval fleurs_tg_test (600 clips — held out, NEVER train on it)
Audio audio_part00..02.tar (≤45 GB shards) — extract ALL into one dir: for t in audio_part*.tar; do tar -xf "$t"; done
Manifests train_rel.jsonl / val_rel.jsonl / fleurs_tg_test_rel.jsonl (relative paths; schema audio_filepath/text/duration)

Preprocessing: transcript dedup, FLEURS-test 8-gram contamination guard, duration 0.4–30 s, resampled to 16 kHz mono.

Public but gated (manual approval).

Licenses: constituent audio is provided under CC-BY-4.0 and compatible terms; credit to the original corpus authors. Underlying audio copyrights remain with their original holders.

Downloads last month
21