SpeechR
收藏资源简介:
SpeechR是一个用于评估大型音频语言模型在语音场景中进行推理能力的统一基准。它从精心策划的文本推理任务生成语音数据,允许对语音内容、韵律和结构进行精确控制,同时保留原始任务的心理意图。SpeechR包含三种推理类型:事实推理、程序推理和规范推理,并发布三个版本:选择题版、生成版和声学特征版,以支持广泛的评估需求。
SpeechR is a unified benchmark for evaluating the reasoning capabilities of large audio language models in speech scenarios. It generates speech data from carefully curated textual reasoning tasks, enabling precise control over speech content, prosody and structure while retaining the psychological intent of the original tasks. SpeechR includes three types of reasoning: factual reasoning, procedural reasoning and normative reasoning, and is released in three versions: multiple-choice version, generative version and acoustic feature version, to support a wide range of evaluation requirements.




