KVoiceBench

KVoiceBench is a Korean spoken question answering (SpokenQA) benchmark derived from VoiceBench. It is part of a Korean speech benchmark suite for evaluating SpeechLMs together with KOpenAudioBench and KMMAU.

Dataset Description

KVoiceBench contains 7,306 Korean spoken QA samples across 9 subsets. It evaluates multiple-choice, binary reasoning, short-answer, open-ended generation, instruction following, and safety behavior under Korean speech input.

KVoiceBench is not the result of naive translation alone. It is constructed through a human-agent SpokenQA benchmark transfer framework:

1. Ground-truth correction audits deterministic source samples before transfer so source-label errors do not propagate into Korean.
2. Hypertranslation transfers source-language questions into Korean using a rulebook that handles direct adaptation, equivalent replacement, deletion, and Korean-specific redesign.
3. Speech-friendly normalization converts written-only forms such as option labels, numbers, units, symbols, abbreviations, and mixed Korean-English expressions into pronounceable Korean text.
4. TTS synthesis generates Korean speech from normalized text using Korean reference voices, followed by ASR-based quality checking and re-synthesis when needed.

During transfer, English-specific task designs are adapted rather than literally translated. For example, because Korean has no direct equivalent to English adjective-ordering constraints, BBH hyperbaton questions are replaced with Korean grammar judgments over particles and endings. Case-sensitive instructions are removed, and letter-frequency constraints are mapped to Korean letters.

Subsets

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

Config	Paper Subset	Samples	Task Type
openbookqa-test	KOpenBookQA	445	Multiple-choice QA
mmsu	KMMSU	3,006	Multiple-choice QA
bbh-test	KBBH	750	Binary reasoning / grammar judgment
sd-qa	KSD-QA	533	Short-answer QA
alpacaeval_full-test	KAlpacaEval	614	Open-ended generation
commoneval-test	KCommonEval	194	Open-ended generation
wildvoice-test	KWildVoice	972	Open-ended generation
ifeval-test	KIFEval	279	Instruction following
advbench-test	KAdvBench	513	Safety / harmful-request refusal

For KVoiceBench, 7,306 of 7,783 source samples are retained after ground-truth correction, hypertranslation, and Korean-specific redesign.

Features

• id (string): Unique sample identifier
• audio (Audio): Synthesized Korean speech audio
• transcription (string): Korean text transcription of the audio
• answer (string): Reference answer, if applicable
• capability (string): Evaluated capability category

Usage

from datasets import load_dataset

# Load a specific subset
ds = load_dataset("anonyemnlpauthor18/KVoiceBench", "bbh-test", split="test")
print(ds[0])

# 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("anonyemnlpauthor18/KVoiceBench", subset, split="test")
    print(f"{subset}: {len(ds)} samples")

Citation

@article{korean_speech_benchmarks2026,
  title={KVoiceBench, KOpenAudioBench, and KMMAU: Agent-Driven Korean Speech Benchmarks for Evaluating SpeechLMs},
  author={Anonymous Authors},
  year={2026}
}

License

This dataset is released under the Apache License 2.0.

© 2026 Anonymous Authors