anonyemnlpauthor18/KVoiceBench

KVoiceBench是一个韩语口语问答基准，源自VoiceBench，是用于评估SpeechLMs的韩语语音基准套件的一部分，与KOpenAudioBench和KMMAU一起。它包含7,306个韩语口语问答样本，覆盖9个子集，评估多项选择、二元推理、简短回答、开放生成、指令遵循和安全行为。数据集构建通过人类代理SpokenQA基准转移框架实现，包括真实值校正、超翻译、语音友好归一化和TTS合成，确保韩语特定适应性和质量。

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. It contains 7,306 Korean spoken QA samples across 9 subsets, evaluating multiple-choice, binary reasoning, short-answer, open-ended generation, instruction following, and safety behavior under Korean speech input. The dataset is constructed through a human-agent SpokenQA benchmark transfer framework involving ground-truth correction, hypertranslation, speech-friendly normalization, and TTS synthesis to ensure Korean-specific adaptation and quality.