Towards Using Virtual Acoustics for Evaluating Spatial Ecoacoustic Monitoring Technologies - Data

This database contains the raw data and certain outputs used for the project 'Towards Using Virtual Acoustics for Evaluating Spatial Ecoacoustic Monitoring Technologies'. In this work, we developed an ambisonic Virtual Sound Environment (VSE) for simulating real natural soundscapes to evaluate spatial PAM technologies in a more controlled and repeatable manner.

We set three objectives to validate this approach:

(O1) to determine whether the VSE could replicate natural soundscapes well enough to be a test environment;

(O2) to pilot the VSE as a test environment for Passive Acoustic Monitoring (PAM) hardware; and, 

(O3) to pilot the VSE as a test platform for PAM software.

To meet these objectives, we used a recently-developed, open source six-microphone field recorder to capture recordings of six field sites and their VSE-based simulations. Sites were based at the Imperial College Silwood Park Campus (Ascot, UK). For O1, we compared field and VSE recordings using a typical suite of ecoacoustic metrics. For O2, we used the VSE to explore how orientation impacts the performance of the six-microphone array. We extended the suite of metrics from O1 to compare VSE recordings from this array at various pitch angles: vertical (as in the field), 45° pitch, and horizontal. For O3, we investigate how BirdNET and HARKBird, software for classifying and localising avian calls, respectively, perform on bird calls added to the VSE-replicated soundscapes. We compare adding calls by encoding to the ambisonics domain and by playback from individual loudspeakers. 

The data is organised as follows: 

• '6mic Audio O1O2': contains the six-channel field and VSE-based recordings of all six sites. Recordings are approximately 10 minutes long. Note that there are three VSE-based recording per site, for each recording orientation used in the VSE (vertical, 45°, and horizontal). Several of our analyses used low-passed versions of these recordings (4 kHz cutoff and 12 dB roll-off); note that the recordings provided here are raw and therefore not low-passed. File names indicate whether the recording was performed in the field or VSE ('Field' vs 'ReRec') and the orientation of the 6mic array for the latter ('V', '45', or 'H'). The final number in each filename indicates the Site the recording corresponds to (1-6).
• '6mic Audio O3': contains the six-channel VSE recordings of five of the original sites with additional avian calls located at certain moments in space and time in each recording. 10 bird calls were added to the soundscapes, each one in its own soundscape recording (each of the five sites' soundscapes was therefore used twice). Four methods of adding the bird calls were trialled, hence there are 40 files in this folder – 10 for each of: ambisonic encoding, playback from individual loudspeakers, playback from individual loudspeakers with no reverb, and playback from individual loudspeakers with no reverb or background soundscape (i.e., 'soloed'). Our analyses focussed on comparing the first of these methods. The filenames are structured as follows: "sXsYbirdNameEmbeddingMethod', where X corresponds to the site the background soundscape was recorded at, Y indicates the position of the sinusoidal sweep used to create simulated reverb of the bird call (1-4 for 0°, 90°, 180°, or 270° around the device used to capture the ambisonic soundscape recordings), 'birdName' is the common name of the added species, and 'EmbeddingMethod' is either 'Ambi' (ambisonic encoding), 'LSPK' (playback from an individual loudspeaker), 'LSPK-NR' (playback from an individual loudspeaker with no reverb), 'Solo' (soloed playback from an individual loudspeaker). Note that due to issues of spatial aliasing, we low-passed the 'Ambi' and 'LSPK' recordings in our analyses (as the two main embedding methods compared) using a 4 kHz cutoff and 12 dB roll-off. However, again the raw (not low-passed) audio is shared here.
• 'Acoustic Indices': contains the CSV files with matrices of the Acoustic Indices extracted from the first channel of the low-passed '6mic Audio O1O2' recordings. Each column is for a different index, rows are the values over time (indices were extracted on 30 s windows). We extracted the following 7 common acoustic indices: Acoustic Complexity Index (ACI), Acoustic Diversity Index (ADI), Acoustic Evenness (AEve), Bioacoustic Index (Bio), Acoustic Entropy (H), Median of the Amplitude Envelope (M), and Normalised Difference Soundscape Index (NDSI). Filenames indicate the field site, whether the indices are for a field or VSE ('Lab') recording, and the orientation of the recording for the latter ('Vert', '45', or 'H'). The 'LP' suffix denotes that these indices were extracted from a low-passed version of the 6mic O1O2 recordings (4 kHz cutoff and 12 dB roll-off).
• 'BirdNET O1O2 Outputs': CSV files generated from avian call classifier BirdNET (using the Winows GUI version) using the first channel ('Mic 1') of the '6mic Audio O1O2' recordings.
• 'BirdNET O3 Outputs': outputs of BirdNET on the '6mic Audio O3' recordings. Here, rather than the raw CSV files generated by BirdNET, the BirdNET results have been filtered to just those during the added bird calls' start and end times, and have been compiled into two CSV files, one for recordings of VSE-based soundscapes with bird calls added by ambisonic encoding and another for calls added by individual loudspeaker playback. These CSV files also contain columns to specify added birds' site, sweep (used to generate reverberation, see above), azimuth and elevation.
• 'HARKBird O1O2 Outputs': outputs of avian call localisation tool HARKBird on the '6mic Audio O1O2' recordings. Note that HARKBird outputs a folder with additional results for each file, however, only the CSV files presented here. These contain the times and estimated azimuth angles of bird calls and were the only HARKBird ouput used for subsequent analyses. Filenames indicate whether recordings are from the field ('Field') or VSE ('ReRec'), and those for the latter also indicate the recording orientation ('Vert', 45' 'H'). The final number in each filename indicates the site the recording corresponds to.
• 'HARKBird O3 Outputs': CSV files of HARKBird outputs (as described above) for the '6mic Audio O3' recordings. Filenames are based on those for this set of recordings (see above); names that contain 'LP' were low-passed (with a 4 kHz cutoff frequency and 12 dB roll-off) prior to passing through HARKBird.
• 'Manual Labels O2': manual labels of audible bird calls' species in the omnidirectional (first) channel of the Zylia ZM-1 ambisonic recordings (see below) of sites 2 and 3. This data was used to calculate the precision and recall of BirdNET's outputs on the 6mic O1O2 recordings for these sites.
• 'VGGish Features': contains the 128-dimension feature embedding of the pre-trained VGGish convolutional neural network extracted from the '6mic Audio O1O2' recordings. Filenames indicate the site, whether the recording was made in the field or VSE ('Lab'), and the recording's orientation. Note again that recordings were low-passed (4 kHz cutoff and 12 dB roll-off) prior to the feature extraction, hence the 'LP' suffix.
• 'Zylia Recordings': approximately 10 minute field recordings of the 6 study sites captured with the 19-microphone ZYLIA 'ZM-1' recorder. These recordings have been converted to third order ambisonic 'b-format' (16 channels) using Furse-Malham channel ordering and SN3D normalisation. This was achieved with the 'Zylia Ambisonics Converter' software. These third-order ambisonic recordings were used to replicate the six sites' soundscapes in the VSE.

The accompanying code for this dataset has been submitted via ScholarOne with the manuscript for peer-review.