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