ohun: an R package for diagnosing and optimizing automatic sound event detection

Abstract

Animal acoustic signals are widely used in diverse research areas due to the relative ease with which sounds can be registered across a wide range of taxonomic groups and research settings. However, bioacoustics research can quickly generate large data sets, which might prove challenging to analyze promptly. Although many tools are available for the automated detection of sounds, choosing the right approach can be difficult only a few tools provide a framework for evaluating detection performance. Here we present ohun, an R package intended to facilitate automated sound detection. ohun provides functions to diagnose and optimize detection routines and compare performance among different detection approaches. The package uses reference annotations containing the time position of target sounds in a training data set to evaluate detection routines performance using common signal detection theory indices. This can be done both with routine outputs imported from other software and detections run within the package. The package also provides functions to organize acoustic data sets in a format amenable to detection analyses. ohun also includes energy-based and template-based detection methods, two commonly used automatic approaches in bioacoustic research. We show how ohun automatically can be used to detect vocal signals with case studies of adult male zebra finch (Taenopygia gutata) songs and Spix's disc-winged bat (Thyroptera tricolor) ultrasonic social calls. We also include examples of how to evaluate the detection performance of ohun and external software. Finally, we provide some general suggestions to improve detection performance.

Evidence for maintenance of key components of vocal learning in ageing budgerigars despite diminished affiliative social interaction

“…Sound files were first downsampled from 44.1 kHz to 19 kHz to improve computational efficiency. Contact calls were selected from original recordings in R v. 4.0.5 [ 38 ] by first using the ‘optimize_energy_detector’ function in the package ohun (v. 0.0.1; [ 39 ]) to apply optimized thresholds for sensitive detection of calls based on amplitude, frequency range, and duration parameters. Spectrograms were generated for a subset of these detections using the ‘spectrograms’ function in warbleR (v. 1.1.27; [ 40 ]) and then visually sorted as either ‘signal’ (contact calls) or ‘noise’ (cage noise, feather ruffling and other vocalization types).…”

“…In addition to contact calls, budgerigars also produce a complex, highly variable and multi-syllabic song called warble, which contains elements that resemble contact calls in structure but differ in behavioural context [ 45 ]. Due to this high acoustic similarity, the random forest model included these contact call-like warble elements in its prediction of calls, requiring a quality control step in which annotations from predicted calls (formatted using the R package Rraven ; [ 46 ]) were visually inspected in Raven Pro v. 1.6 [ 47 ] so that warble elements could be removed, yielding a total of 49 573 confirmed contact calls out of the 150 534 predicted calls.…”

“…The first two dimensions were plotted to generate an acoustic trait space wherein acoustically similar calls appear closer together. We then used the R package PhenotypeSpace (v. 0.1.0; [49]) to quantify acoustic areas to assess the effect of age on an individual's acoustic repertoire over time. To characterize the acoustic space of each individual per audio recording block, we computed kernel density areas for calls subset by individual identity and time, which quantifies irregular areas with higher precision [50].…”

The evolution of sexually dimorphic traits in ecological gradients: an interplay between natural and sexual selection in hummingbirds

“…The acoustic features quantify the distribution of energy in the time and frequency domain, and the variation in dominant frequency across time (see electronic supplementary material, methods for details). Song acoustic space was quantified as the minimum spanning tree connecting all elements of the song in the overall acoustic space, using the R package PhenotypeSpace [65]. Four parameters describing song complexity were calculated for each song: element types, acoustic space, element transition diversity and between-song variation (electronic supplementary material, methods).…”

When night becomes day: Artificial light at night alters insect behavior under semi-natural conditions

“…(2021), using "R", version v.3.4.1. (52), the "Rraven" open source package(53), and RavenPro1.5(54).All bioacoustic data, aka stridulation events, were manually validated to avoid false positives/negatives and to eliminate other acoustic sources in the semi-natural conditions, such as bird vocalizations. Only stridulation activity patterns containing at least five consecutive days and nights of behavioral data were used for further analysis.Data processing and statistical analyses were conducted in Python version 3.PyCharm, JetBrains), SPSS version 21 (IBM Corp. Armonk, NY, USA), and Prism 8 (GraphPad Software, San Diego, California USA).…”