KVoiceBench

Paper: KVoiceBench, KOpenAudioBench, and KMMAU: Agent-Driven Korean Speech Benchmarks for Evaluating SpeechLMs

KVoiceBench is a Korean spoken question answering benchmark for evaluating speech models, built upon VoiceBench.

Dataset Description

As there is no existing benchmark for evaluating Korean spoken question answering capabilities, we construct KVoiceBench by building upon the English benchmark VoiceBench.

Construction Process

Following the paper's SpokenQA transfer framework, KVoiceBench was built by transferring VoiceBench into Korean. Source samples with deterministic answers were first audited through a reviewer/meta-reviewer LLM process; confirmed label errors were corrected and invalid cases were excluded. The retained items were then hypertranslated with a human-agent rulebook. A separate human-agent normalization loop audited the hypertranslated files, organized likely TTS failure cases into a normalization rulebook, and produced speech-friendly Korean before Qwen3-TTS synthesis with Korean reference voices and WER-based resynthesis checks.

Subsets

The dataset consists of 7,306 samples across 9 subsets:

Subset	Samples	Task Type
KOpenBookQA	445	Multiple-choice QA
KMMSU	3,006	Multiple-choice QA
KBBH	750	Binary reasoning
KSD-QA	533	Short-answer QA
KAlpacaEval	614	Open-ended
KCommonEval	194	Open-ended
KWildVoice	972	Open-ended
KIFEval	279	Instruction following
KAdvBench	513	Safety

Features

• id (string): Unique sample identifier
• audio (Audio): Synthesized Korean speech audio
• transcription (string): Korean text transcription of the audio
• answer (string): Reference answer (empty string for open-ended/safety subsets)
• capability (string): Evaluated capability category

Usage

from datasets import load_dataset

# Load a specific subset
ds = load_dataset("KRAFTON/KVoiceBench", "bbh-test", split="test")
print(ds[0])
# {'id': 'bbh_navigate_147', 'audio': {'path': ..., 'array': array([...]), 'sampling_rate': 24000},
#  'transcription': '...', 'answer': '...', 'capability': 'context_related/multi_hop_reasoning'}

# Load all subsets
subsets = [
    "advbench-test", "alpacaeval_full-test", "bbh-test", "commoneval-test",
    "ifeval-test", "mmsu", "openbookqa-test", "sd-qa", "wildvoice-test"
]
for subset in subsets:
    ds = load_dataset("KRAFTON/KVoiceBench", subset, split="test")
    print(f"{subset}: {len(ds)} samples")

Citation

@misc{kim2026kvoicebenchkopenaudiobenchkmmau,
  title={KVoiceBench, KOpenAudioBench, and KMMAU: Agent-Driven Korean Speech Benchmarks for Evaluating SpeechLMs},
  author={Kim, Haechan and Chung, Seungjun and Park, Inkyu and Lee, Jihoo and Lee, Jonghyun},
  year={2026},
  eprint={2605.27984},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/pdf/2605.27984v1}
}

License

This dataset is released under the Apache License 2.0.

© 2026 KRAFTON