Auditory perception of natural soundscapes: Hearing biodiversity – HEARBIODIV

To address these issues, a massive acoustic database is built up including a variety of terrestrial biomes. Acoustic recordings are analyzed using dedicated signal-processing tools to identify “statistics” distinguishing terrestrial biomes, and biophony from geophony (geophysical sounds such as rain or wind). Computational models of the human auditory system associated with automatic classifiers are developed to predict human behavior in response to natural or artificial soundscapes. Psychophysical paradigms are designed to assess the ability of human observers to perceive these natural or artificial soundscapes, and empirical data are compared to simulated data to evaluate the optimality of human observers in these tasks.

(i) We have recorde or retrieved natural soundscapes across a variety (N=13) of terrestrial biomes located on five continents (America, Europe, Africa, Asia, Oceania). Recordings were obtained for at least two contrasting seasons. These biomes correspond to: temperate deciduous and/or coniferous forests (n=3), boreal forest (n=1), tropical forests (n=3), sub-tropical forest (n=1), savannah (n=1), desert (n=1), grassland (n=1), meadow (n=1), and chaparral (n=1). Overall, this massive acoustic database (wav format) representing a variety of terrestrial biomes with minimal human impact has a size of 12 TB.
(ii) A critical review of the eco-acoustic literature conducted by a member of ISYEB and ENS-LSP led us to clarify the ‘soundscape’ concept, and develop a theoretical framework to explain how natural soundscapes are produced. This framework will guide the investigations of the present project. Two computational models of the human auditory system including a cochlear filterbank and a central modulation filterbank have been developed and refined during the first year.
(iii) Several experimental paradigms have been designed to evaluate human capacities to i) discriminate natural soundscapes and ii) discriminate artificial assemblages differing in terms of number of individual vocalizations of birds or bird species. A forced-choice oddity paradigm was chosen for all experimental designs. Although being difficult, this forced-choice paradigm discourages bias and allows to compute percent correct and sensitivity (d’) scores for each task. A questionnaire has been designed to assess lifelong exposure of our participants to natural soundscapes, and thus evaluate the importance of experience and amount of exposure in soundscape perception. Several vocoders have been developed to process natural soundscapes in order to evaluate the importance of spectral and temporal cues in soundscape perception. Data collection and analysis has been conducted on several groups of normal-hearing adult participants recruited via the RISC platform.

Participants’ performance was well above chance for unprocessed stimuli, indicating high sensitivity for natural soundscape discrimination. Training did not affect performance. Additional results indicate that temporal cues play only a minor role. Instead, listeners appeared to base their decisions primarily on gross spectral cues related to biological sound sources and habitat acoustics. Simulations of an optimized convolutional neural network using spectro-temporal information extracted at the output of a model of human perception confirm that listeners exclude the available temporal information in the discrimination of natural soundscapes.

In everyday life, humans engage with various “soundscapes” corresponding to the acoustical patterns produced by biological, geophysical and anthrophonic sound sources. This program aims to assess, clarify and model the sensory mechanisms allowing humans to discriminate auditory textures evoked by soundscapes recorded in nature reserves; in particular, we will assess to which extent and how humans discriminate textures associated with changes in biodiversity. The project will be executed in three stages: (i) Identification of relevant texture statistics for biodiversity assessment; ii) Selection of soundscape recordings in protected areas and analysis; (iii) psychophysical assessment of texture discrimination capacities for humans, focusing on discrimination of soundscapes differing in terms of species richness. Deviations of listeners from performance of automatic classifiers will help us probing information processing and decision statistics used by humans for biodiversity assessment.