text stringlengths 3 716 | audio_bytes listlengths 14.4k 986k | audio_size int64 16.6k 640k |
|---|---|---|
"дар ҷопон император дар атрофи дарбори императорӣ бах(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,5,-88,0,18,-28,3,-24,0,-16,0,3,-64,0,63,-51,-43,-93,38,110,93,6,-(...TRUNCATED) | 245,760 |
"об аз болои сарбанд ба қитъаи бараш 100 фут рехта мешава(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,1,-30,0,18,50,3,-24,0,-16,0,1,-107,0,-109,-13,72,-24,-117,-38,17,(...TRUNCATED) | 103,680 |
"собиқ корманди пулиси филлипинӣ сайёҳони ҳонконгиро (...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,4,21,0,19,-21,3,-24,0,-16,0,3,105,-64,-125,-121,-96,125,-97,-53,5(...TRUNCATED) | 223,680 |
"агар шумо дар фасли зимистон ба арктика ё антарктида т(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,0,11,0,14,124,3,-24,0,-16,0,4,7,64,-22,-23,55,102,-38,-39,57,-116(...TRUNCATED) | 264,000 |
"занон бояд дарк намоянд ки фарқиятҳои фарҳангӣ метаво(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,3,-51,0,17,102,3,-24,0,-16,0,4,89,-64,107,-89,36,-84,-125,-50,13,(...TRUNCATED) | 285,120 |
"қудрати зафаровари ӯ ба ҳама аз подшоҳ то ба фуқаро за(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,1,-54,0,18,42,3,-24,0,-16,0,1,-29,-64,23,31,-97,92,65,-83,-75,-11(...TRUNCATED) | 123,840 |
"дар якҷоягӣ бо нисбатан дастнорас будани он тимбукту (...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,0,-120,0,17,-24,3,-24,0,-16,0,3,113,64,79,-116,-29,6,20,99,78,-11(...TRUNCATED) | 225,600 |
"қабл аз тағйир додани моддаҳо розигии якдилонаи ҳамаи(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,0,11,0,16,7,3,-24,0,-16,0,3,-94,0,-114,115,-106,102,71,-34,-94,88(...TRUNCATED) | 238,080 |
"барои мисол омӯзиш ва иҷтимоишавӣ ҳамчун унсурҳои анг(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,0,11,0,17,-13,3,-24,0,-16,0,3,19,-128,-3,81,101,69,63,-49,-49,-12(...TRUNCATED) | 201,600 |
"пулро танҳо дар бонки ягона дар ҷазираҳо ки дар стэнли(...TRUNCATED) | [102,76,97,67,0,0,0,34,16,0,16,0,0,0,11,0,14,25,3,-24,0,-16,0,2,-11,-128,-34,21,-102,-38,17,-78,-107(...TRUNCATED) | 193,920 |
tajik-asr-corpus-v3
1,071 hours of Tajik ASR training data: 41 Tajik YouTube channels (~1,059 h, machine-labeled)
plus FLEURS tg_tj (11.8 h, gold). This is the dataset behind
Peacockery/omni-ctc-300m-tajik
(16.9% WER on FLEURS test, 37.6% on held-out conversational speech).
Layout
Hive-partitioned parquet under version=0/corpus=<source>/split=<split>/language=tgk_Cyrl/.
Each row holds text (the normalized label), audio_bytes (16 kHz mono FLAC as an int8 list),
and audio_size (sample count). language_distribution_0.tsv lists hours per corpus.
Test split
split=test holds two partitions: FLEURS test (gold labels, exported unfiltered), and a
conversational held-out set — 1,625 clips from 157 whole videos carved out before training, so
no clip from those videos appears in train. The conversational references are machine labels
that passed the same agreement bar as training data, which makes that partition a relative
benchmark; FLEURS is the absolute anchor.
How the labels were made
FLEURS keeps its original transcripts. YouTube audio carries machine labels from an ElevenLabs Scribe ensemble, kept only where independent passes agree (WER ≤ 0.35 between passes), then language-gated (Tajik Cyrillic vs Russian/Persian) and normalized for CTC training. Machine labels are unverified by native speakers.
Sources
FLEURS: CC-BY-4.0, google/fleurs. YouTube: public channels, labels produced by this project; per-clip channel and video identifiers are in the corpus partition names.
- Downloads last month
- 172