开源时间：2025年1月

       MusicEval 数据集是一个生成式音乐评估数据集，总时长 16.62 小时，包含 2748 个单声道音乐片段，以及 14 位音乐专家给出的 13740 条评分。这些片段由 21 个不同的系统（涵盖 31 个模型）响应 261 条提示词生成。261条文本提示词包含80条人工撰写，20条来自MusicCaps数据集，这二者提供给模型生成音乐；其余161条仅对应提供样例的系统所生成的音乐片段。为保证数据一致性，所有音乐片段均为为 16kHz 单声道格式。每个音乐片段均由来自音乐学院的 5 位音乐专家进行背靠背评估，专家基于整体音乐印象与文本提示的契合度两项标准进行打分。这两个维度分别聚焦音乐的质量本身，以及音乐与给定文本提示的一致性，为音乐生成模型的综合评估提供了关键基准。

The MusicEval dataset is the first generative music assessment dataset. The dataset contains 2,748 music clips generated by 31 prevalent and advanced TTM (Text-to-Music) models in response to 261 text prompts, along with 13,740 ratings collected from 14 music experts. The total duration is 16.62 hours.

The MusicEval dataset collects 261 text prompts in total. Among these, 80 manually written prompts and 20 prompts selected from the MusicCaps dataset are used for the open-access models to generate music, while the remaining 161 descriptions only correspond to music clips from the demo-only system.

All music clips are 16khz mono audio and each music clip is evaluated by 5 raters on two dimensions: overall musical impression and alignment with the text prompt, which respectively emphasize the importance of both the quality of the generated music and its consistency with the given text prompt.

论 文

数据下载

查看样例

MusicEval 生成式音乐评分数据集

A Generative Music Dataset with Expert Ratings for Automatic Text-to-Music Evaluation

* This dataset was jointly developed and constructed by the HLT Laboratory of the College of Computer Science at Nankai University and AISHELL.

 License: CC BY NC 4.0