Sunocaps: A Novel Dataset of Text-Prompt Based Ai-Generated Music with Emotion Annotations

The SunoCaps dataset aims to provide an innovative contribution to music data. Expert description of human-made musical pieces, from the widely used MusicCaps dataset, are used as prompts forgenerating complete songs for this dataset. This Automatic Music Generation is done with the state25 of-the-a...

Descripción completa

Detalles Bibliográficos
Autores: Civit, Miguel, Drai-Zerbib, Véronique, Lizcano, David, Escalona, María José
Tipo de recurso: artículo
Fecha de publicación:2024
País:España
Institución:Universidad a Distancia de Madrid (UDIMA)
Repositorio:udiMundus. Repositorio Institucional de la Universidad a Distancia de Madrid
OAI Identifier:oai:udimundus.udima.es:20.500.12226/2169
Acceso en línea:http://hdl.handle.net/20.500.12226/2169
http://dx.doi.org/10.2139/ssrn.4832849
Access Level:acceso abierto
Palabra clave:Data
Automatic Music Generation
Emotion feature
Artificial Intelligence
Prompt algnment
Generative AI
Descripción
Sumario:The SunoCaps dataset aims to provide an innovative contribution to music data. Expert description of human-made musical pieces, from the widely used MusicCaps dataset, are used as prompts forgenerating complete songs for this dataset. This Automatic Music Generation is done with the state25 of-the-art Suno generator of audio-based music. A subset of 64 pieces from MusicCaps is currently included, with a total of 256 generated entries. This total stems from generating four different variations for each human piece; two versions based on the original caption and two versions based on the original aspect description. As an AI-generated music dataset, SunoCaps also includes expert-based information on prompt alignment, with the main differences between prompt and final generation annotated. Furthermore, annotations describing the main discrete emotions induced by the piece. This dataset can have an array of implementations, such as creating and improving music generation validation tools, training systems for multi-layered architectures and the optimization of music emotion estimation systems.