TurkASR-Bench combines seven Turkish speech sources that keep their own licenses (see the license column and LICENSE.md). By requesting access you agree to use the data for evaluation and research, to respect the license of each source (the Khan Academy subset is non-commercial, CC BY-NC-SA 3.0; the VoxForge subset is GPL-3.0), and not to attempt to identify the speakers.

TurkASR-Bench

A balanced Turkish speech-recognition test set: 743 utterances from each of 7 sources (5,200 in total, 14.7 hours). It is the test set behind the Turkish ASR Leaderboard, where 14 open models are ranked with WER, CER and KDHO, a root- and affix-aware error rate for Turkish.

Equal weight per source means that no single kind of speech (read news, lectures, spontaneous dialogue, …) dominates the average. All audio is standardized to 16 kHz mono FLAC (lossless), the format the models consume.

Quick start

from datasets import load_dataset

ds = load_dataset("yagmurtuncer/turkasr-bench", split="test")                  # all 7 sources
cv = load_dataset("yagmurtuncer/turkasr-bench", "common_voice", split="test")  # one source

x = ds[0]
x["audio"]["array"], x["audio"]["sampling_rate"], x["text"]

Score a model the way the leaderboard does (pip install kdho jiwer):

import kdho
refs = ds["text"]
hyps = [my_asr(a["array"]) for a in ds["audio"]]
kdho.score(refs, hyps, verbalize_numbers=True)   # {'kdho': ..., 'wer': ..., 'cer': ...}, lower is better

Sources

Config	Utterances	Hours	Speech	License
`common_voice`	743	0.57	Crowdsourced read sentences (Common Voice 17)	CC0-1.0
`fleurs`	743	2.61	Read Wikipedia sentences (Google FLEURS)	CC-BY-4.0
`khan_academy`	743	2.13	Educational lectures	CC-BY-NC-SA-3.0 (non-commercial)
`mediaspeech`	743	2.95	Broadcast media (MediaSpeech)	CC-BY-4.0
`real_turnturk`	742	0.65	Spontaneous two-person dialogue	CC-BY-4.0
`voxforge`	743	1.29	Read prompts, volunteers (VoxForge)	GPL-3.0
`youtube`	743	4.52	YouTube video transcripts, varied topics	CC-BY-4.0

Utterances were drawn from each source with a fixed seed (42). One Real-TurnTurk file with empty audio was removed.

Fields

Field	Content
`id`	Utterance id (same as in the leaderboard outputs)
`source`	Config name
`audio`	16 kHz mono audio
`text`	Reference transcript as published by the source
`text_normalized`	Benchmark normalization: NFC, Turkish lowercasing (I→ı, İ→i), punctuation removed (word-internal apostrophes kept)
`text_normalized_numbers`	Same, with digits spelled out in Turkish (48 → kırk sekiz), the leaderboard's main setting
`duration_s`	Duration in seconds
`kdho_split`	`dev` / `test` split used while designing the KDHO metric; report final numbers on all rows or on `test`
`license`	License of the source
`suspected_reference_typo`	Likely spelling errors in the reference found by model agreement (e.g. `kesinlike → kesinlikle`); not manually verified
`origin`	Where the utterance comes from in the original source (JSON: repo, row index, file or segment id)

Known limitations

References are used as published by each source. An automatic audit flagged 175 utterances (191 words) whose reference probably contains a spelling error, mostly in VoxForge; they are marked in suspected_reference_typo and were not corrected. Correcting all of them would change WER by at most 0.2 points.
Some references write numbers as words and some as digits; use text_normalized_numbers to compare fairly.
The XLS-R models on the leaderboard were fine-tuned on Common Voice; their Common Voice scores may benefit from speaker overlap.

Citation

@misc{turkasrbench2026,
  title  = {TurkASR-Bench and KDHO: A Balanced Turkish ASR Benchmark with a Root- and Affix-Aware Error Rate},
  author = {Tuncer, Nur Yağmur},
  year   = {2026},
  url    = {https://huggingface.co/datasets/yagmurtuncer/turkasr-bench}
}

Please also cite the original sources: Common Voice (Ardila et al., 2020), FLEURS (Conneau et al., 2023), MediaSpeech (Kolobov et al., 2021), VoxForge, Khan Academy and the YouTube transcript collection.

Downloads last month: 32

Total file size:

938 MB

Access TurkASR-Bench