The VU Sound Corpus: Adding More Fine-grained Annotations to the Freesound Database

The VU Sound Corpus Adding More Fine-grained Annotations to the Freesound Database Emiel van Miltenburg, Benjamin Timmermans, Lora Aroyo The Network Institute Vrije Universiteit Amsterdam {emiel.van.miltenburg, b.timmermans, lora.aroyo}@vu.nl Abstract This paper presents a collection of annotations (tags or keywords) for a set of 2,133 environmental sounds taken from the Freesound database (www.freesound.org). The annotations are acquired through an open-ended crowd-labeling task, in which participants were asked to provide keywords for each of three sounds. The main goal of this study is to find out (i) whether it’s feasible to collect keywords for a large collection of sounds through crowdsourcing, and (ii) how people talk about sounds, and what information they can infer from hearing a sound in isolation. Our main finding is that it is not only feasible to perform crowd-labeling for a large collection of sounds, it is also very useful to highlight different aspects of the sounds that authors may fail to mention. Our data is freely available, and can be used to ground semantic models, improve search in audio databases, and to study the language of sound. Keywords: Sounds, sound terms, Crowdsourcing, Corpus, Onomatopoeia 1. Introduction Recent years have seen a growing interest in annotated col- lections of perceptual stimuli. Most of the attention has been directed towards either image labeling (Deng et al., 2009) or image description (Young et al., 2014; Lin et al., 2014). This paper presents a collection of annotations (tags or keywords) for a set of environmental sounds taken from the Freesound database.1 The annotations are acquired through an open-ended crowd-labeling task. The main goal of this study is to find out how people talk about sounds, and what information they can infer from hearing a sound in isolation. Our data can be used to ground semantic mod- els, improve search in audio databases, and to study the language of sound.2 1.1. Freesound and Other Sound Collections The Freesound database (Font et al., 2013) is an open col- laborative database of almost 300 000 sounds released un- der Creative Commons licenses. Each sound is provided by its author with a description and a set of keywords. Though there are some guidelines for the descriptions,3 au- thors are free to add whatever keywords they want. There are also some more standardized datasets, but all of these are much smaller. Hocking et al. (2013) provide a compre- hensive overview of environmental sound norming studies, of which Saygin et al. (2005) present the largest dataset with 236 annotated sounds. All studies use sounds with a duration between ten seconds and less than a second. 1.2. Grounded Semantic Models In distributional semantics, the meaning of a word is a func- tion of the contexts in which that word occurs (Turney et al., 2010). Up until recently, ‘context’ was simply taken to mean ‘surrounding words’. Recent work in multimodal distributional semantics challenges this notion because text 1www.freesound.org 2The code and data is available at: https://github. com/CrowdTruth/vu-sound-corpus 3See http://freesound.org/help/faq/ corpora form “an extremely impoverished basis compared to the rich …

Content truncated to 500 words