With growing interest in ecologically valid stimuli and the proliferation of recognition and classification algorithms, environmental sounds are increasingly a topic of interest and study. However, there is a lack of consensus as to which sound categories should be selected, and which exemplars should represent these categories. Indeed, many existing datasets include only one exemplar per sound class. This is not characteristic of the actual acoustic environment and can lead to brittle and unrepresentative computational solutions and experimental results. Importantly, the existing literature provides relatively little information about how human listeners perceive the recognizability, similarity and representativeness of acoustically and perceptually varying sounds. Here, we introduce the Environmental Sounds (EnviSounds) dataset. It provides data from over 1000 human listeners performing different tasks (recognition, similarity judgement, goodness of fit) on 53 commonly encountered environmental sound categories of human, natural and man-made sounds, each represented by 10 different exemplars. We provide behavioural indices of within-class similarity and goodness of exemplar for all samples. Furthermore, we include normative data for recognition and identification latency, accuracy, and imageability, enabling researchers to select items for experiments based on pre-defined criteria across these dimensions. With this dataset, environmental sounds can be considered at various levels of complexity: sounds produced by the same sources varying in their acoustics (within-class acoustic variability) and acoustically similar sounds produced by different sources (between-category confusion). The dataset can be used for specifying training sets with individual sound sources, defining the categorical boundaries, or providing additional variables to building prediction models.
Kachlicka et al. (Thu,) studied this question.