rustam1221/uzbek-asr-benchmark-spontaneous

Uzbek Spontaneous Speech ASR Benchmark

A 311-clip, 1.98-hour test set for Uzbek speech recognition, built from spontaneous YouTube speech: podcasts, interviews and multi-speaker conversation with overlapping turns, fillers, code-switching into Russian, and dialect spelling. Every transcript that an automatic difficulty check flagged as possibly wrong was corrected by hand — 118 of the 311 — and the protocol below says exactly which ones and why. Version 3 (2026-10-28) adds a second, listening-based pass: every place where our best model disagreed with the reference was checked by ear, and 144 clips were corrected.

It exists because the public Uzbek ASR benchmarks do not measure the thing we ship. FLEURS is read speech, and models score two to three times better on it than on real conversation. On this benchmark the same GigaAM-large baseline that scores 6.7% WER on FLEURS scores 14.4%. Anything tuned against FLEURS alone is tuned against the easy case.

Clips	311
Audio	1.98 h
Version	v3 (2026-10-28)
Reference words	12,593
Clip length	3.2–30.0 s (median 28.5 s)
Human-verified clips	118 reviewed word by word (v2) + 167 adjudicated by ear (v3)
Domain	spontaneous YouTube speech (podcast, interview, vlog)
Script	Latin Uzbek, punctuated
Audio included	no — reconstructed from the upstream source (see below)

Why the transcripts can be trusted

The source corpus (Abduqayum/Uzbek-STT-Dataset-780h) ships Gemini-generated transcripts. Machine transcripts cannot be used to score machine transcripts, so every clip went through a verification pass:

1. Sampling. 315 clips drawn with a fixed seed. Clips containing music, Cyrillic script, or digits were excluded before sampling — those are text normalization problems, not recognition problems, and they would have made the numbers about the wrong thing.
2. Difficulty scoring. Each clip got an asr_wer field: the disagreement between the Gemini transcript and GigaAM-large. High disagreement means at least one of the two is wrong.
3. Targeted human review. A 7-clip random spot check of the low-disagreement band came back 100% correct, so manual review was spent where it pays: all 122 clips with asr_wer >= 0.18 were reviewed word by word in a purpose-built annotation UI. 118 were corrected or confirmed, 4 were discarded as unusable. The remaining 193 clips were auto-accepted.
4. Result (v2). 311 clips: 118 human-verified, 193 auto-accepted behind a measured error rate.
5. Listening adjudication (v3). Every substitution where uzasr-large-v0.3.0 disagreed with the reference (814 spans) was first sorted by an LLM from text alone: reference wrong, model wrong, spelling variant, dialect form, or unsure. A human then listened to every "reference wrong" and "unsure" span and chose: reference, model, neither (typing what was actually said), or inaudible. 374 of 450 spans were resolved; the 76 unresolved "unsure" spans were left untouched. The reference changes only where a human confirmed it by ear — 144 clips. The LLM agreed with the listener on 84 of its 92 "reference wrong" calls; the other 8 are why text-only judgement was not trusted on its own.

This is a deliberate trade. Full manual transcription of 2 hours costs roughly 20 annotator-hours; this protocol spent about 4 and put them where the uncertainty was.

Files

File	What it is
final_etalon.jsonl	the benchmark (v3) — 311 rows, id, source_row, duration, text, reviewed, asr_wer
final_etalon_v2.jsonl	the previous version, for reproducing older numbers
changes_v3.jsonl	before/after text of the 144 clips corrected in v3
adjudication_v3.jsonl	every listened span: reference words, model words, LLM pre-sort, human decision, typed text
etalon.jsonl	all 315 sampled rows before discards, with the original Gemini text
corrections.jsonl	the 122 human review decisions (ok / discard + corrected text)
source_map.json	id → upstream parquet file + row, for rebuilding the audio
exclude_rows.json	the same rows keyed by parquet file, to remove them from training data
predictions/*.jsonl	reference/hypothesis pairs for every model in the table below
scoring.py	normalization + WER/CER, standalone
reconstruct_benchmark_audio.py	rebuilds the audio from upstream

etalon.jsonl, corrections.jsonl and the *_v3 files are kept so both verification passes are auditable: anyone can diff the machine transcript against the human decision.

Getting the audio

Audio is not redistributed here. Every row carries its upstream location, and the helper script fetches only the parquet row groups that hold benchmark clips rather than the full 780-hour corpus:

pip install pyarrow huggingface_hub
python reconstruct_benchmark_audio.py ./benchmark_audio

Output is <id>.wav, mono 16 kHz, matching final_etalon.jsonl by id.

Results

Corpus-level WER/CER with jiwer. Text on both sides is lowercased, mapped to canonical Uzbek apostrophes (ʻ/ʼ) and stripped of punctuation before scoring.

Model	Params	WER	CER	Canonical WER	WER on v2
large_agree_600m + KenLM	600M	8.1%	2.5%	7.4%	10.8%
large_full_600m + KenLM	600M	9.4%	2.8%	8.7%	11.5%
large_full_600m	600M	10.4%	2.8%	9.8%	12.2%
base_agree_220m + KenLM	220M	9.4%	3.0%	8.7%	11.3%
gemini_full_220m + KenLM	220M	10.5%	3.3%	9.9%	12.2%
gemini_full_220m	220M	12.4%	3.3%	11.8%	13.9%
GigaAM-large — untuned	600M	14.4%	5.3%	13.2%	16.1%
whisper-large-v3-turbo-uzbek	809M	38.6%	18.5%	38.2%	39.0%†

† Re-run on 2026-10-28 with the same code, weights and library versions. The August 2026 run, whose predictions were not kept, gave 39.6% / 19.4% on v2. The 0.6-point drift comes from the runtime environment, not from the references. The untuned GigaAM-large reproduced to within 0.05 points.

Older models, scored on v2 only (their per-clip predictions were not kept, so they cannot be re-scored; compare them only with the "WER on v2" column):

Model	Params	WER (v2)	CER (v2)
gemini_bench_220m	220M	14.0%	3.8%
punct_220m	220M	15.0%	4.1%
base_220m	220M	15.5%	4.5%
GigaAM — untuned	220M	17.1%	5.4%

For reference, the same untuned GigaAM-large scores 6.7% WER on FLEURS. The gap between 6.7% and 14.4% (16.1% on v2) is the entire reason this benchmark exists.

The shipped configurations are large_agree_600m + KenLM, tag uzasr-large-v0.3.0, and base_agree_220m + KenLM, tag uzasr-base-v0.3.0. The KenLM text excludes every sentence that shares a 10-word sequence with any benchmark reference, plus ±25 neighbouring rows, and the decoder weights were chosen by 2-fold cross-validation on this set (held-out WER 11.11% and 12.26% on v2).

Correction (2026-10-17). KenLM rows were earlier reported as 10.2% / 10.8%. 62 of the 311 clips (same recordings re-published in other Uzbek YouTube datasets) had leaked into the LM text through sources the row-level exclusion did not cover; the LM scored 5.8% on them. Anyone training on public Uzbek YouTube corpora should check for the same overlap — n-gram matching against the references here finds it in minutes. Their per-clip predictions are in predictions/, so any row of this table can be recomputed or disputed.

Canonical scoring. Uzbek clitics (da, ku, chi, mi, yu, ya, ta) are written both joined and separated in real text, and neither is wrong. Scoring them as errors measures orthographic convention, not recognition. The canonical variant joins clitics on both sides before scoring; it is worth about 0.6 WER points and is the number to compare against.

Reproducing a score

scoring.py in this repository implements the normalization and both WER variants, with no other dependency than jiwer:

import json
from huggingface_hub import hf_hub_download
from scoring import score

rows = [
    json.loads(line)
    for line in open(
        hf_hub_download(
            "rustam1221/uzbek-asr-benchmark-spontaneous",
            "final_etalon.jsonl",
            repo_type="dataset",
        )
    )
]

references = [row["text"] for row in rows]
hypotheses = [your_model.transcribe(f"benchmark_audio/{row['id']}.wav") for row in rows]
print(score(references, hypotheses))
# {'wer': ..., 'cer': ..., 'wer_canonical': ..., 'cer_canonical': ..., 'wer_gate': ...}

Already have predictions on disk? python scoring.py predictions/large_full_600m.jsonl prints all five numbers for any JSONL with ref and hyp fields. wer_gate additionally drops filler sounds (a, mm, eh) on both sides; it is our internal release gate, not a headline number.

Known limits

• Long clips. Median length is 28.5 s, near the source corpus cap. Errors concentrate in the longest multi-speaker clips, so a model that segments well before recognizing has an advantage this benchmark does not isolate.
• Auto-accepted tail. The 193 auto-accepted clips inherit any Gemini error that GigaAM-large happened to agree with. The spot check and the v3 listening pass reduce this, they do not eliminate it.
• Adjudication bias (v3). The listening pass looked only where uzasr-large-v0.3.0 disagreed with the reference, so v3 fixes the errors that model exposed, not ones other models would expose. That model gained 2.6 WER points from v3, the others 1.4–2.1; the ranking did not change. Future models get the same pass on their own disagreements; for uzasr-base-v0.3.0 that pass has not run yet, so its v3 number is, if anything, pessimistic.
• Single source corpus. All clips come from one upstream YouTube collection, so channel and topic diversity is whatever that collection has.
• Size. 1.98 h and 12.4k words: differences under roughly half a WER point are not meaningful.

Licensing

The annotations in this repository (corrections, review decisions, row mapping) are released for research and evaluation use. The underlying audio belongs to its original YouTube publishers and is not redistributed here — it is fetched from the upstream dataset, whose terms apply to that audio.

Citation

@misc{uzbek_spontaneous_asr_benchmark,
  title  = {Uzbek Spontaneous Speech ASR Benchmark},
  author = {Nuriddinov, Rustamjon},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rustam1221/uzbek-asr-benchmark-spontaneous}
}