Data from: Animal-borne acoustic recordings of African elephant vocalizations from Samburu National Reserve, Kenya

African elephants are a species of conservation concern with complex vocal behavior. Annotated acoustic recordings of elephants can be used both for studying vocal behavior, including call rates and types, and developing acoustic detection and classification models. We present an annotated dataset of African elephant vocalizations from an affordable acoustic recorder (AudioMoth) integrated into a GPS collar deployed onto a wild female African elephant in Kenya. A total of 62 hours of animal-borne acoustic data were recorded and 1,123 vocalizations were identified including 706 high-quality rumbles that were further used in acoustic feature analysis. The average group call rate was 18.1 calls/hr (range: 0 - 104), with a change point analysis identifying three periods of declining call rate after the collaring event. Our findings offer insight into the vocal behavior of African elephants, and the data provided can be used in the development of acoustic detection models. # Data from: Animal-borne acoustic recordings of African elephant vocalizations from Samburu National Reserve, Kenya This dataset contains audio recordings collected using an AudioMoth recorder deployed onto a GPS collar on an African elephant named Arden in Samburu, Kenya, in June 2025. ## Folder structure **audio.zip** (3.73 GB) - zipped folder containing 62 one-hour .wav files recorded at a 16 kHz sample rate. The start time of each file (local time, East Africa Time) is encoded in the filename, e.g. `Arden_20250606_170125.WAV`. The beginning of the filename indicates the name of the individual (Arden). **clips.zip** (120.8 MB) - zipped folder containing 1,123 .wav clips at the original sample rate (16 kHz), each containing one manually identified rumble extracted from the raw audio with a 0.2 s buffer added before and after the call. Each clip filename includes the source recording, the selection number of the call within that recording, the call type, and a quality indicator. **labels.zip** (52.8 KB) - contains 62 .txt annotation Raven files, one per raw audio file, with matching filenames. Three recordings had no vocalizations identified during manual review; empty annotation files are included for these so that the file counts between the raw audio and label folders match. **model.zip** (403.75 KB) - contains the automated rumble detection model output, run separately on 8 kHz audio and 4 kHz audio. Each of the two `4kHz` and `8kHz` subfolders contains: * a combined **annotation** file with one row per detection across all 62 recordings * a **summary** file with one row per recording, giving the file duration and total detection counts * a **per_file_annotations** subfolder, containing the per-recording split of the combined annotation file (one .txt file per source recording, matching the manual label filenames). These annotations can be directly opened and viewed in Raven. ## Manual annotation file format Each manual label file is a tab-delimited text file with one row per annotated call and the following columns: * **Selection** - ID number assigned to the call. Numbers may skip if a selection was removed during review. * **View** - spectrogram view label, constant across all files (Spectrogram 1). * **Channel** - recording channel, always 1 since all recordings are mono. * **Begin Time (s) / End Time (s)** - start and end time of the call, in seconds from the start of the recording. * **Low Freq (Hz) / High Freq (Hz)** - lower and upper frequency bounds of the annotation box. * **call_type** - RUM (rumble), ROAR, or ROAR-RUM (a call containing both). * **quality** - overall quality rating of the call, on a 0-4 scale (see Table 2 in the associated manuscript for the full description). * **overlap** - whether the call overlapped in time with another vocalization (Y/N). * **earflap** - presence of acoustic interference from ear flapping, coded 0-2 (see Table 2 in the associated manuscript for the full description). All annotations were made manually in Raven Pro. ## Model output format ### Annotation files `model/4kHz/audiomoth_downsampled_4kHz_annotation.txt` (1136 detections, 4 kHz audio) and `model/8kHz/audiomoth_downsampled_8kHz_annotation.txt` (1,395 detections, 8 kHz audio) contain the automated call detections from the trained model across all 62 recordings combined. These files use the same tab-delimited Raven selection table format as the manual labels, with additional columns generated by the detection pipeline: * **Selection** - ID number assigned to the detection. * **View** - spectrogram view label, constant across all rows. * **Channel** - recording channel; always 1 (mono). * **Event** - class label assigned by the model (Rumble for all detections in this dataset). * **Begin Time (s) / End Time (s)** - start and end time of the detection, in seconds from the start of the full concatenated recording stream (not the individual file). * **Low Freq (Hz) / High Freq (Hz)** - lower and upper frequency bounds of the detection. * **Begin Path** - full file path to the source audio file at the time the model was run. * **File Offset (s)** - start time of the detection, in seconds from the start of the individual audio file named in Begin File. * **Begin File** - filename of the audio file containing the start of the detection. * **Site** - recording site/device name. For this dataset, this is the name of the collared elephant (Arden). * **Begin Hour** - local hour of day (0-23) when the detection begins. * **File Start Date / Begin Date** - calendar date (local time) of the file start and of the detection. * **Score** - model confidence score for the detection. * **Count, Measurable, Harmonics, Ambiguous, Notes, Analyst, Rand, Deployment, Sound Problems, Call Criteria, Disk** - fields carried over from the Raven annotation template; left blank in the automated output and not used for the model results. ### Summary files `model/4kHz/audiomoth_downsampled_4kHz_summary.txt` and `model/8kHz/audiomoth_downsampled_8kHz_summary.txt` each contain one row per source recording, with the following tab-delimited columns: * **Begin Path** - full file path to the source audio file at the time the model was run. * **File Duration (s)** - length of the recording in seconds. * **Rumble Count** - number of rumble detections in that recording. * **Gunshot Count** - number of gunshot detections in that recording (0 throughout this dataset; the model was not applied to a gunshot-annotated training set for this deployment). ### Per-file annotations `model/4kHz/per_file_annotations/` and `model/8kHz/per_file_annotations/` each contain one .txt file per source recording (e.g. `Arden_20250606_130550_annotations.txt`), giving the subset of the combined annotation file corresponding to that recording, in the same column format described above. Unlike the manual `labels` folder, recordings with zero model detections do not have a corresponding file in these subfolders (59 files for 4 kHz; 61 files for 8 kHz).