Daana-dev: Kyrgyz read-speech ASR benchmark (public dev split)

Anonymous copy for peer review. Daana is a Kyrgyz read-speech benchmark of 1,500 sentences, each read by both of two readers (3,000 clips, 7.28 h). The sentences are split into a hidden test set (981 sentences) and this public dev set (519 sentences, 1,038 clips, 2.51 h). References are corrected to what each reader said; the paper describes how and estimates the error that remains.

Contents

path	content
wav/	dev clips, 16 kHz mono 16-bit WAV: the exact files every system read
manifest_dev.jsonl	one line per dev clip: references (reference_eval is scored), reader (SPK-F / SPK-M), register, length class, contamination tier, Track A flag, SHA-256 of the audio
split.csv	sentence_id,split for all 1,500 sentences
decision_tables/dev/	every reference decision for dev clips, in full
decision_tables/test/	the same tables for test clips without any text (ids, classes, votes, verdicts)
results/*/ours/hypotheses_dev.jsonl	raw hypotheses of each system on dev
results/leaderboard/	test and dev aggregates as reported in the paper
supplement/	text-free per-clip error counts on the hidden test, code that recomputes every test interval and pairwise test, and SHA-256 commitments to the test manifest and hypotheses
tools/, runners/, normalization/, systems/	evaluation harness, normaliser, system registry
paper/generated/	every number and table in the paper, as generated

Organisation names in corpus labels and paths are replaced by anon in this review copy.

Quick start

pip install -r requirements.txt
python3 tools/score.py --manifest manifest_dev.jsonl \
    --hyp results/gigaam_large_ctc/ours/hypotheses_dev.jsonl \
    --out dev_scores --track a
cd supplement && python3 recompute.py     # test aggregates from text-free counts

The first command reproduces the dev column of the results table (Track A: digit-free clips of at most 30 s). To evaluate a new system, write a hypotheses.jsonl with the runner in runners/ and score it the same way.

Readers and consent

Two adult volunteer readers, not paid, each read every sentence. Both gave written informed consent covering public release of the recordings. Released identifiers are pseudonymous (SPK-F, SPK-M); age is given as a band and origin as a region. A voice remains a biometric trait.

Licence and rights

We hold no rights to the prompt texts. They come from several sources (a CC-BY-NC-4.0 compilation of Kyrgyz news articles, TV drama transcripts, literary prose, TV popular science and folk tales); rights remain with their holders, and the texts are distributed sentence by sentence for evaluation only. Our licence, CC-BY-NC-4.0, covers the recordings and annotations (corrected references, decision tables). In practice the dev package as a whole may be used only for non-commercial evaluation. Code: Apache-2.0 (LICENSE).

Do not train or fine-tune on this split if you report results on Daana-test: dev shares the test's two readers. Systems trained on dev are listed separately.

Hidden test

The test audio and references are not distributed. Open-weight models are evaluated on request under the paper's protocol, and researchers can obtain the test on request for evaluation only, under an agreement that excludes training and redistribution. Contact details will be given in the published paper.

Citation

Anonymous submission; citation withheld for review.