MusicLM by Google

MusicCaps is a unique dataset designed for music understanding and description tasks. It consists of short music clips, each labeled with detailed descriptions written by musicians. These descriptions include a list of specific musical aspects, like "pop, tinny wide hi hats, mellow piano melody," and a free-text caption that captures the overall sound and mood of the clip. The dataset is sourced from AudioSet and is divided into training and evaluation sets, allowing for development and testing of music-related AI models.

Highlights:

• Expert-written descriptions: Each music clip is meticulously described by musicians, providing rich and accurate information about its sound.
• Detailed aspect lists: Precise musical aspects are identified and labeled for each clip, enabling detailed analysis of musical elements.
• Free-text captions: Descriptive captions offer a comprehensive understanding of the music's style, mood, and instrumentation.
• Ready-to-use dataset: The dataset is curated and split for training and evaluation, facilitating the development of AI models for music understanding.

Key Features:

• Music clips with detailed annotations: Each clip is accompanied by aspect lists and free-text captions.
• Sourced from AudioSet: The dataset leverages the vast collection of labeled audio clips in the AudioSet database.
• Creative Commons license: The dataset is licensed under Creative Commons, allowing for various research and development purposes.
• Metadata-rich: Clips are labeled with metadata such as YouTube video ID, start and end time, AudioSet labels, author ID, and more.