
source audio	321 audiobook files, 9.49 GB
transcribed	300 books
gold segments	34,114 (152.94 h)
held in review	415 (2.74 h)
rejected	7
script	30,716 Latin / 19 Cyrillic segments gold

Layout

audio/ — the raw full-length recordings, unmodified.
manifest/gold.jsonl — segments that pass every structural check. Training set.
manifest/review.jsonl — segments held back with the reason recorded. Kept, not discarded.
manifest/reject.jsonl — segments failing a hard check.
manifest/SUMMARY.json — counts, hours, and the gate thresholds actually used.

One manifest record:

{"id": "uzab_s2_1f3c9a2b41_00007", "audio_hf": "audio/<file>.mp3",
 "start": 41.22, "end": 52.86, "duration": 11.64,
 "text": "...", "text_norm": "...", "script": "latn",
 "avg_logprob": -0.114, "no_speech_prob": 0.01, "tier": "gold"}

audio_hf is a path inside this repo, so a segment can be cut directly from audio/. text is the transcript as decoded; text_norm is case-folded and punctuation-stripped with the okina (U+02BB) preserved as a letter and apostrophe variants folded onto it, which is the form to score against.

How the transcripts were made

Transcribed with an in-house Uzbek/Central-Asian fine-tune of Whisper (whisper-ca e2, 10.88 WER on FLEURS uz), faster-whisper + Silero VAD, beam 5, chunk_length=20.

Segments then pass a structural gate: duration 1–30 s, 4–24 characters/second, avg_logprob ≥ −1.0, no_speech_prob ≤ 0.6, compression ratio ≤ 2.6, and ≥60% of letters in one Uzbek script (Latin or Cyrillic both accepted — text that is predominantly neither, usually Russian or English, is held).

Label trust is lower than for a dual-model corpus, and we state that rather than paper over it. Our Tajik pipeline gated on agreement between two different architectures. No comparable second Uzbek model exists in-house — the next best is 53.54 WER against Whisper's 10.88, and a peer that much weaker measures its own errors, not Whisper's. So for Uzbek the structural gate is the primary filter, not a tiebreaker. These are machine labels: good enough to train on, not human-verified.

Known gaps

Three source downloads are corrupt (two truncated at power-of-two boundaries, one empty) and are excluded. A handful of source files were still transferring when this version was published and are added in a later commit. Some gold-adjacent segments exceeded the 30 s window and sit in review; re-segmenting them would need a narrower window than the one measured to help, so they were left in review rather than forced through.

Licence

The recordings are third-party audiobook material collected for research. No ownership is claimed over the audio. Access is manually gated; request access stating your intended use.

Downloads last month: -

Total file size:

9.19 GB