rustam1221/uzbek-asr-train-manifests

Uzbek ASR Training Manifests

The exact training, validation and test splits behind rustam1221/uzbek-asr-gigaam: 974 hours of Uzbek speech drawn from seven public corpora, filtered, text-normalized, and split by speaker.

No audio is copied. Each row is a pointer — a parquet file plus a row index in the upstream dataset — and the training dataloader decodes the audio when the batch is built. That keeps the whole corpus definition at 200 MB instead of roughly a terabyte of duplicated audio, and it makes the manifests a reviewable artifact: a git diff shows exactly which rows entered a run.

Splits

Split	Rows	Hours	Punctuated hours
train	620,342	916.3	795.1
val	28,543	36.5	30.9
test	13,439	21.9	19.7

Punctuated hours matter because punctuation is trained as a second phase over the subset that still carries . , ? ! in its original text.

Sources

Source	Train hours	Rows	What it is	Upstream license
uzbekvoice	490.4	416,141	crowd-read sentences	open
usc	97.0	97,730	Uzbek Speech Corpus, read speech	open
yt_it	91.7	13,103	YouTube IT talks, Russian/English code-switching	Apache-2.0
common_voice	86.9	70,597	Mozilla Common Voice uz, validated split only	CC0
yt_news	86.5	12,399	YouTube news reads	Apache-2.0
yt_podcasts	56.2	8,105	YouTube podcasts, spontaneous	Apache-2.0
fleurs	7.5	2,267	FLEURS uz, kept as a comparable benchmark	CC-BY

usc carries no punctuation in its source text and is therefore excluded from the punctuation phase.

Not in this snapshot

The models were later extended with yt_gemini — roughly 660 accepted hours from Abduqayum/Uzbek-STT-Dataset-780h, spontaneous YouTube speech with LLM-generated transcripts. That source is part of the training recipe and of every gemini_* and large_* checkpoint, but its manifest snapshot is not published here; the published snapshot is the 916-hour base mixture, which is what every non-gemini checkpoint saw.

The 315 benchmark rows are removed from yt_gemini before training, keyed by exclude_rows.json in uzbek-asr-benchmark-spontaneous. Leakage was checked after the exclusion: zero benchmark rows reached the training set.

Row format

{"parquet": "uzbekvoice_filtered/data/train-00003-of-00042.parquet", "row": 1517,
 "duration": 4.82, "text": "bugun havo juda issiq boʻldi",
 "text_raw": "Bugun havo juda issiq boʻldi.", "source": "uzbekvoice", "split": "train"}

Field	Meaning
parquet + row	location of the audio in the upstream dataset
audio_filepath	used instead of parquet/row where audio is a local file (FLEURS, extracted from tar)
duration	seconds
text	normalized: Latin, lowercase, canonical apostrophes (ʻ/ʼ), no punctuation
text_raw	original text with punctuation, for the punctuation phase
source	one of the seven source names above
split	train / val / test

Filtering

Applied uniformly to every source, adapted from the uzbek-asr-curated-701h recipe:

• rows containing digits are dropped — spoken form is a text-frontend problem, and leaving them in teaches the model to guess
• rows with characters outside the Uzbek Latin alphabet are dropped
• duration outside 0.5–30 s
• more than 18 characters per second, or outside 0.3–2.67 words per second — these catch truncated audio and runaway transcripts

Splitting

Deterministic hash of the speaker id, 96/2/2. Speaker-disjoint splits matter here: uzbekvoice is 45% of the corpus and heavily repeats speakers, so a random row split would leak voices into validation and flatter the numbers. Sources without a speaker column are hashed on the row identity. FLEURS keeps its own dev/test split untouched so its numbers stay comparable with published results.

Integrity

test_manifest.jsonl   b89219fdb4f5a02f4e9b7cfab5e38972c750c215ddef9c745eb193f9145a46d2
train_manifest.jsonl  c1b32999ef1a2a842d7c7b2bb94899ab45ffb470ef0a1600245ba2e9444f388b
val_manifest.jsonl    403ebb22bf88ed9c75b87134b33ee627f20a4bca912ac893633893deb71fd8e2

Every training run records the config, the git commit and these manifest hashes, so a checkpoint can be traced back to the exact corpus definition it saw.

Licensing

Each source keeps its own upstream license; the combination is not redistributable under a single license, which is why this repository ships pointers and not audio. Check the upstream terms before using any source commercially.