DAACI-VoDAn Dataset

DAACI-VoDAn is a dataset for vocal detection that comprises manually-generated vocal activity annotations for 706 full-length music tracks, as well as their associated metadata including song title, artist, and YouTube URL. This data repository contains the metadata file and the annotations as CSV files, and a pickle file with the proposed data partitions: train (80%), validation (10%), and test (10%). The music tracks in DAACI-VoDAn were selected to cover a broad variety w.r.t. genre, era, and instrumentation, and are intended to be representative of the music universe available through commercial catalogues and streaming services. <strong>Citing, license, and using the dataset</strong> This dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License and it is intended to be used only for research purposes. DAACI-VoDAn is presented in the following paper, accepted for publication at EUSIPCO (European Signal Processing Conference) 2023. Please, cite this paper when using the dataset: Helena Cuesta, Nadine Kroher, Aggelos Pikrakis, Stojan Djordjevic. <strong>DAACI-VoDAn: Improving Vocal Detection with New Data and Methods</strong>. To appear in Proceedings of the European Signal Processing Conference (EUSIPCO). 2023. <strong>Annotation format</strong> Vocal regions are provided as CSV files with a unique track ID as filename (matching the IDs specified in the metadata file). They were created manually by a team of four annotators following a peer-reviewing strategy, i.e., each annotation was created by one annotator and reviewed by another. Each row in the CSV files represents a vocal region given by the tuple [start_time, duration] in seconds.<br> <br> The task of annotating vocal segments deals with numerous ambiguous scenarios. To provide consistent and extendable annotations, we defined an annotation strategy in the form of a set of rules: We defined a minimum inter-region distance of 200 ms. Two vocal regions which are at least 200 ms apart are annotated as separate regions. Shorter vocal rests are not considered and a single long region is annotated instead. We annotated heavily processed (e.g., chorus or distortion effects) singing voice sections as vocals, as long as they are still recognisable as originating from a human voice. We extended vocal sections to include reverb and delay “tails" as long as they are audible. Similarly, we considered inhalations or “breathing" sounds, e.g., at the beginning of a vocal phrase, as part of vocal section. We considered segments containing speech as vocal sections but excluded whistling. Vocal ensembles and background vocals are also labelled as vocals.