Calibrated microphone array recordings reveal that a gleaning bat emits low-intensity echolocation calls even in open-space habitat

ABSTRACT Echolocating bats use ultrasound for orientation and prey capture in darkness. Ultrasound is strongly attenuated in air. Consequently, aerial-hawking bats generally emit very intense echolocation calls to maximize detection range. However, call levels vary more than tenfold (>20 dB) between species and are tightly linked to the foraging strategy. The brown long-eared bat (Plecotus auritus) is a primarily gleaning, low-amplitude species that may occasionally hawk airborne prey. We used state-of-the-art calibrated acoustic 3D-localization and automated call analysis to measure P. auritus’ source levels. Plecotus auritus emits echolocation calls of low amplitude (92 dB rmsSPL re. 20 µPa at 10 cm) even while flying in open-space. While P. auritus thus probably benefits from delayed evasive manoeuvres of eared insects, we propose that low-amplitude echolocation did not evolve as an adaptive countermeasure, but is limited by morphological constraints.

1 kHz, all corrected for the frequency response of the loudspeaker) from a loudspeaker (custom-built Polaroid speaker).The loudspeaker was placed at 50 cm distance to a ¼-inch measuring microphone with flat frequency response (40BF, with preamplifier 26AC and power module 12AA; GRAS Sound and Vibration, Holte, Denmark) and to the microphone used for recording the bat calls in the field (FG-O, Avisoft Bioacoustics, Glienicke, Germany).The sensitivity of the measuring microphone was calibrated by recording a 1 kHz tone at 94 dB SPL rms (sound calibrator 4231,Brüel & Kjaer,Naerum,Denmark).

Accuracy of 3D-localization and source-level calculations
We performed two calibration measurements to determine the accuracy of our acoustic tracking system (consisting of our array and the TOADSuite software package) for calculating (i) the 3D-coordinates of bats and (ii) the (apparent) source levels of bat calls.As acoustic signals for calibration, we used two types of echolocation-like signals: a linear downward frequency-modulated sweep (FM,, and a logarithmic downward frequency-modulated sweep (mimicking a FM-QCF call of e.g., pipistrelle bats; 90-25 kHz).
Each signal was generated with three different durations (3, 6, and 9 ms) with a Hann window (to avoid clicks in the loudspeaker) and repeated five times, so that the total playback comprised 30 sweeps, sandwiched between two 100-ms long audible pure tones of 10 kHz (total playback duration 4.8 sec) to indicate the start and end of a playback sequence.The ramp segment of the Hann window is dependent on the total duration of the sound and thus varies with the signal durations.Hence the measured signal duration at -12 dB below peak amplitude will be lower than the specified values.
First, we broadcast this playback from a loudspeaker (custom-built Polaroid speaker) at 3, 4.5, 6, 8.5, and 10 m distance to the microphone array in a large hall.In addition to the onaxis recording (i.e., 0° azimuth and elevation), we also rotated the array relative to the loudspeaker to record the playback from 30° and 60° off-axis in both elevation (below the array's central axis) and azimuth (to the left of the array's central axis).We measured the distance between the microphone and the loudspeaker with a laser range-finder (GLM 50C, Bosch, Germany) for every distance and angle adjustment.In total, we thus recorded 1,350 sweeps (30 sweeps per broadcast (2 sweeps × 3 durations × 5 repetitions) × 5 distances × 3 directions in azimuth × 3 directions in elevation).
Second, we analysed the array recordings of the calibration playback with the TOADSuite in the same way as the bat calls recorded in the field (see main text).In short, we band-passfiltered the recordings (18-90 kHz), automatically detected the broadcast sweeps, determined the time-of-arrival-differences of each sweep between the microphones by cross-correlation, calculated each sweep's three-dimensional position, compensated the sweep's recorded waveform for the frequency-and direction-dependent microphone sensitivity and for the distance-dependent geometric and atmospheric attenuation, and then calculated each sweep's source level from the compensated waveform (both as peak-equivalent SPL and RMS-SPL).
This resulted in the 3D-position and source levels of the calibration playback as calculated by the TOADSuite, which we compared in the next steps to the real 3D-positions (as measured by the laser range finder, see above) and the real source level (as calibrated with a measuring microphone, see next step).
Third, we measured the real source level of our playbacks with a calibrated ¼-inch measuring microphone with flat frequency response (40BF, with preamplifier 26AC and power module 12AA; GRAS Sound and Vibration, Holte, Denmark) in an echo-attenuated room.We placed the Polaroid loudspeaker at 50 cm to the measuring microphone and recorded the playback consisting of 30 sweeps.We used custom-written scripts (provided online) in R version 3.6.1 (R Core Team, 2019) to analyse the calibration recording in the same way as the TOADSuite to obtain the real source levels of our playback.Specifically, we band-pass-filtered the recording (18-90 kHz), detected the 30 sweeps, and calculated for each sweep its peakamplitude from the envelope of the waveform and the RMS-value of the waveform within the -12 dB call duration criterion.By comparing these amplitude measurements to the recording of a calibration tone (94 dB SPL rms, 1 kHz; sound calibrator 4231, Brüel & Kjaer, Naerum, Denmark), we calculated the real source level of both sweeps, both as peak-equivalent value and RMS value, expressed in dB SPL re.20 µPa at 10 cm distance.

3D Location:
We determined the accuracy of our acoustic tracking system to calculate the threedimensional coordinates and the source level of a sound source.From the reconstructed 3Dposition of the loudspeaker, we calculated the reconstructed distance between loudspeaker Journal of Experimental Biology: doi:10.1242/jeb.245801:Supplementary information Journal of Experimental Biology • Supplementary information and central microphone.To quantify the spatial accuracy of our acoustic tracking system, we compared the reconstructed distance to the real distance that we measured with the laser range finder.We express the spatial error between the reconstructed and real distance as the percentage of the absolute difference between the reconstructed and real distance, relative to the real distance: (1) spatial error (%) = (reconstructed_distance -real_distance) / real_distance * 100.
The effects of the different variables (Distance to the array, Azimuth, Elevation, call duration, call design) on the error in sound source localisation and source level were estimated using a Gaussian Generalized Linear Model (GLM) with identity link function in a Bayesian framework (package rstanrm v. 2.21.3, in R v. 4.2.2).Four chains of 5000 iteration each were run (warmup 2500 iteration, thinning rate 1) using the default normal non-informative prior.
We examined the quality of model fit following (Gabry, 2022) (package bayesplot v. 1.10.1)by examining the MCMC trace, ensuring the Rhats and effective sample size were acceptable, and ensuring that there was minimum autocorrelation within each chain.
The overall mean localization error was close to Zero (-0.5% of the real distance), with a standard-deviation of 4.1% and a range from -9.4 -10.9% (Fig. S1).This small mean error indicates that the tracking system does not systematically over-or underestimate the real position.Calculating a smoothed trajectory of the noisy raw data will thus likely result in a good estimate of the real flight trajectory, and excluding raw data points that are more than ~10% off the smoothed trajectory appears to be justified.

Journal of Experimental Biology • Supplementary information
In addition to calculating the error for the overall distance, we performed the same calculations separately for each of the three spatial dimensions (X, Y, Z), i.e., we express the spatial error for each dimension by comparing the reconstructed distance to the real distance in that dimension.The X-and Z-axes form the plane of the microphone array, with X being the horizontal and Z the vertical axis, and the Y-axis is the direction away from the array.
Source Level: To quantify the acoustic accuracy of our acoustic tracking system in measuring source levels, we compared the reconstructed source level (SL) to the real source level obtained from the measuring microphone.We express the accuracy of source level reconstruction as the difference between reconstructed and real source level: (2) SL error (dB) = reconstructed_SL -real_SL.
We used the same statistical method to analyse the effect of call type and distance to the microphone as explained int eh previous section.The mean source levels of our playback signals (relative to 20 µPa and at 10 cm distance to the loudspeaker) were 115.1 dB peSPL (111.8 dB SPL RMS) for the FM signal, and 110.1 dB peSPL (106.1 dB SPL RMS) for FM-QCF signals (±0.1 dB SD in all cases for the 15 sweeps per call type).The source levels calculated by the TOADSuite were slightly lower than these values, resulting in a mean (±SD) source level error of -1.1±1.7 dB for peSPL and -4.2±1.5 dB for rmsSPL (Fig. S2).
Journal of Experimental Biology: doi:10.1242/jeb.245801:Supplementary information In summary, source level estimates for peak-values are quite accurate (-1 dB) when averaged over sufficient calls, while the error is larger for average (RMS) estimates (-4 dB).Lower signal-to-noise ratios increased the error but did not lead to any systematic bias in source level measurements (Fig.

S1 D).
Potential reasons for the larger error in RMS-estimates might be fluctuations in the call envelope caused by interference patterns under real-world recording conditions, incomplete compensation of the microphone's frequency response, and other unknown effects.

Call duration and frequency of calls recorded in the field
The call duration and frequency parameters show some variation and a few outliers (Fig. 3C).
Analysing call duration as a function of the signal-to-noise ratio (SNR) showed that the call durations and frequencies are reliable for SNR above ~20 dB.For SNR <20 dB, the call duration and frequency of some (but not all) sweeps increased, yet with no obvious pattern regarding which sweep was affected, and no difference between FM and FM-QCF sweeps (Fig. S1 B,C).One reason for this pattern might be that with decreasing SNR, the call definition threshold (-12 dB re.maximum amplitude) falls below random fluctuations of the envelope and thus includes noise into the call definition, explaining the few very high call durations in Fig. 3. Overall, we did not find any pattern explaining the extreme outliers in duration and frequency (e.g., long-eared bat calls with long duration were not the same as the calls with very high peak frequency or lowest frequency; no pattern in SNR or localization error) and thus included all data into the analysis.

Fig. S1 .
Fig. S1.Accuracy of measuring 3D-positions (A) and call parameters (B-D) with a 60cm star-shaped microphone array and the TOADSuite a) The tracking accuracy is indicated by the distance error, which is the difference between the distance calculated with the TOADSuite and the real distance measured with the laser range finder, expressed as a percentage of the real distance.Data is presented as a function of the distance to the microphone, and separated by elevation (panels from top to bottom) and azimuth (color-coded), for two call types (FM signals, left, and FM-QCF signals, right).Raw data is shown behind boxplots, which show median, quartiles, and 95 th -percentiles.b) Measured call duration as a function of the signal-to-noise-ratio.Signal duration and shape are coded with shades and colours.As comparison, the dashed lines show calculated durations for the playback at 1 m distance and recorded with a measurement microphone and analysed independently in R using the same method implemented in the TOADSuite.Note