QCRI/MenaSpeechBank

MENASpeechBank是一个以中东和北非（MENA）为中心的参考语音库，旨在支持在现实的多轮助手交互设置下训练和评估音频大型语言模型（AudioLLMs）。数据集包含高质量的语音片段和人物角色配置文件，支持多轮对话结构、说话者/方言多样性、基于人物角色的上下文和约束等。具体包括来自124位独特说话者的17,641条语音片段（约26.4小时）和469个人物角色配置文件。数据集适用于音频到文本响应生成、长上下文口语对话记忆评估、跨方言/口音和信道变异性的鲁棒性分析等研究用途，但不建议用于说话者识别或生物特征分析等用途。

MENASpeechBank is a MENA-centric reference voice bank designed to support training and evaluation of AudioLLMs under realistic, multi-turn assistant interaction settings. It includes high-quality utterances and persona profiles, supporting multi-turn structure, speaker/dialect diversity, persona-grounded context and constraints. Specifically, it contains 17,641 utterances (~26.4 hours) from 124 unique speakers and 469 persona profiles. The dataset is intended for research uses such as audio-to-text response generation, long-context spoken dialogue memory evaluation, and robustness analysis across dialect/accent and channel variability, but not recommended for speaker identification or biometric profiling.