Digital Audio Recordings of the Song of Unseen Orthoptera: Guidelines for the Empirical Visual and Aural Species-Level Attribution and Introduction of the “Sandwich Method”
Year: 2026
Type: preprint
Abstract: Field recordists often deal with situations where the presence of a species is documented exclusively by an audio file, without accompanying images nor physical specimens. Machine Learning (ML) classifiers trained on annotated datasets are increasingly popular for the automated identification of insect species from soundscapes or single-species recording. Such promising technology is generally reliable, but cannot correctly classify a species whose time/frequency patterns is not included in the reference datasets. Human experts cannot match the speed of a ML classifier but can provide the flexibility needed to overcome its limitations, by taking advantage of the digital audio software commonplace in the field of bioacoustics and of simple operations such as cut and paste. Such empirical comparisons are complicated by several factors, here individually addressed: guidelines are provided to mitigate the difficulties of specific recognition, both by emphasizing the nature of the obstacles, and by providing step-by-step procedures to overcome them. In light of a growing concern for the lack of verifiability in arthropod research, and a corresponding lack of repeatability in bioacoustic studies, and considering that science is the study of repeatable falsifiable phenomena with a methodology designed to minimize confirmation/cognition biases, the paper introduces the empirical “sandwich method”, that involves two audio files: “the filling”, a time slice of 2 seconds – or one echeme, if its duration exceeds two seconds – cut from the novel, unrecognized recording and inserted in “the bread”, the exemplary reference audio. On-screen view is then zoomed to “the sandwich”, a time window of six seconds (or at least three echemes) centered on the filling. Recognition is based on the similarity of bread and filling as observed side by side first visually (oscillogram), and then aurally (unaided ear comparison). Recognition of the poorly audible sandwiches can be improved by the frequency-lowering effect granted by playing the audio at lower speed (e.g., 1/4 or 1/10 of the natural speed) thanks to the special functions of the audio software. The method requires to save the decisive sandwich as an audio file and as a screenshot, thus ensuring persistence, verifiability, repeatability and possible redaction of the recognition.
Source: Preprints.org
Authors: Cesare Brizio, Filippo Maria Buzzetti
Language: English
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7204834236
