# Context People have been making music for tens of thousands of years [1]. Today, making music is easier and more accessible than ever before. The technological developments of the last few decades allow people to simulate playing every imaginable instrument on their computers. Audio sequencers enable users to arrange their songs on a time line, sample by sample. Digital audio workstations (DAWs) ship with virtual instruments and synthesizers which allow users to virtually play a whole band or orchestra in their bedrooms. One challenge in working with DAWs is organizing samples and recordings in a structured way; so, users can easily access them. In addition to their own recordings, many users download samples. Browsing through sample collections to find the perfect sound is time consuming and may impede the user's creative flow [2]. On top of this, manually naming and tagging recordings is a time-consuming and tedious task, so not many users do [3]. The consequence is that finding the right sound at the right moment becomes a challenging problem [4]. Modeling the relationship between the acoustic content and semantic descriptions of sounds could allow users to retrieve sounds using text queries. This dataset was collected to support research on content-based audio retrieval systems, focused on sounds used in creative context. # Content This dataset was collected from [Freesound](http://freesound.org]) [5] in June 2016. It contains the frame-based MFCCs of about 230,000 sounds and the associated tags. - `sounds.json`: Sound metadata originally downloaded from the Freesound API. This file includes the `id`, associated `tags`, links to `previews`, and links to an `analysis_frames` file, which contains frame-based low-level features, for each sound. - `preprocessed_tags.csv`: Preprocessed tags. Contains only tags which are associated to at least 0.01% of sounds. Moreover, tags were split on hyphens and stemmed. Tags containing numbers and short tags with less than three characters were removed. - `queries.csv`: An aggregated query-log of real user-queries against the Freesound database, collected between May 11 and November 24 in 2016. - `preprocessed_queries.csv` Queries were preprocessed in the same way tags were preprocessed. - `*_mfccs.csv.bz2`: The original MFCCs for each sound, extracted from the URL provided in the `analysis_frames` field of `sounds.json`, split across ten files. - `cb_{512|1024|2048|4096}_sparse.pkl`: Codebook representation of sounds saved as sparse `pd.DataFrame`. The first-order and second-order derivatives of the 13 MFCCs were appended to the MFCC feature vectors of each sound. All frames were clustered using K-Means (Mini-Batch K-Means) to find {512|1024|2048|4096} cluster centers. Each frame was, then, assigned to its closest cluster center and the counts used to represent a sound as a single {512|1024|2048|4096}-dimensional vector. # Acknowledgements Thanks to the Music Technology Group of the *Universitat Pompeu Fabra* in Barcelona for creating and maintaining the Freesound [5] database and for providing the aggregated query-logs. # Inspiration Who can create the best content-based audio retrieval system measured by precision-at-*k* for values of *k* in {1, ..., 20} and mean average precision. # Getting started Here's the accompanying GitHub repository: https://github.com/dschwertfeger/cbar # References [1] N. L. Wallin and B. Merker, The Origins of Music. MIT Press, 2001. [2] M. Csikszentmihalyi, Flow: The Psychology of Optimal Experience. New York: Harper Perennial Modern Classics, 2008. [3] E. Pampalk, A. Rauber, and D. Merkl, "Content-based organization and visualization of music archives