Conversational recommendation has advanced rapidly with large language models (LLMs), yet music remains a uniquely challenging domain in which effective recommendations require reasoning over audio content beyond what text or metadata can capture. We present MusiCRS, the first benchmark for audio-centric conversational recommendation that links authentic user conversations from Reddit with corresponding tracks. MusiCRS includes 477 high-quality conversations spanning diverse genres (classical, hip-hop, electronic, metal, pop, indie, jazz), with 3,589 unique musical entities and audio grounding via YouTube links. MusiCRS supports evaluation under three input modality configurations: audio-only, query-only, and audio+query, allowing systematic comparison of audio-LLMs, retrieval models, and traditional approaches. Our experiments reveal that current systems struggle with cross-modal integration, with optimal performance frequently occurring in single-modality settings rather than multimodal configurations. This highlights fundamental limitations in cross-modal knowledge integration, as models excel at dialogue semantics but struggle when grounding abstract musical concepts in audio. To facilitate progress, we release the MusiCRS dataset (https://huggingface.co/datasets/rohan2810/MusiCRS), evaluation code (https://github.com/rohan2810/musiCRS), and comprehensive baselines.

翻译：随着大语言模型（LLMs）的快速发展，对话推荐系统取得了显著进步，然而音乐领域仍是一个独特的挑战性领域，其中有效的推荐需要对音频内容进行推理，这超出了文本或元数据所能捕捉的范围。我们提出了MusiCRS，这是首个面向音频中心对话推荐的基准测试，它将来自Reddit的真实用户对话与相应的音乐曲目关联起来。MusiCRS包含477个高质量对话，涵盖多种音乐流派（古典、嘻哈、电子、金属、流行、独立、爵士），涉及3,589个独特的音乐实体，并通过YouTube链接提供音频基础。MusiCRS支持在三种输入模态配置下进行评估：仅音频、仅查询以及音频+查询，从而允许对音频-LLMs、检索模型和传统方法进行系统比较。我们的实验表明，当前系统在多模态集成方面存在困难，最优性能往往出现在单模态设置中，而非多模态配置。这凸显了跨模态知识集成的基本局限性，因为模型擅长处理对话语义，但在将抽象音乐概念与音频基础关联时却面临挑战。为促进该领域发展，我们发布了MusiCRS数据集（https://huggingface.co/datasets/rohan2810/MusiCRS）、评估代码（https://github.com/rohan2810/musiCRS）以及全面的基线模型。