A collection of fully-annotated soundscape recordings from the Western Ghats in India

This collection contains 1,115 soundscape recordings with a total duration of more than 79 hours, which have been annotated by expert ornithologists who provided 30,324 bounding box labels for 116 different bird species from the Western Ghats, including numerous threatened or endangered native birds. These data were collected between July 2018 and January 2021 across multiple forest sites, ranging from actively restored to undisturbed rainforest habitats. Parts of the collection were featured as the hidden test set in the BirdCLEF 2024 competition and is intended for evaluating machine learning models for species detection and classification in complex acoustic soundscapes. Data collection Soundscapes in this collection were recorded as part of two projects: a) a long-term ecological monitoring project to study the effects of rainforest restoration on bird communities in the Anamalai Hills of the Western Ghats and b) long-term avian research in the shola sky islands of the Palani hills in the Western Ghats. For the first project, AudioMoth autonomous recording units (ARUs) were deployed at various forest sites (along a gradient of forest regeneration), mounted ~2 meters above ground level, and configured to record short audio segments every 5 minutes over seven consecutive days. Recordings were made between March 2020 and January 2021 (excluding April 2020 due to COVID-19-related restrictions), using a sample rate of 48 kHz and a gain setting of 40 dB. For the Palani hills, we deployed eleven ARUs in different stands of invasive timber plantations and natural forests; nine Song Meter 4 and two Song Meter 2 (Wildlife Acoustics Inc). The ARUs were scheduled to record for 5 minutes every 10 minutes at 16-bit, 44000 kHz in .wav format, across 24 hours from July 2018 to July 2019. All audio was converted to FLAC, and resampled to 32 kHz for this collection. Parts of the collection were featured as the hidden test set in the BirdCLEF 2024 competition. Sampling and annotation protocol This dataset is a subset of recordings from the broader restoration monitoring effort. A team of expert ornithologists reviewed the soundscapes both aurally and using spectrograms (in Raven Pro) to identify vocalizations of all bird species present at each site. Each soundscape was segmented into 10-second clips—the minimum duration required for accurate species-level identification. For each 10-second segment, vocalizations were annotated with time-frequency bounding boxes, and each label was linked to an eBird species code based on the 2021 Clements taxonomy. All faint, ambiguous, or overlapping calls that could not be reliably identified were excluded. In total, the dataset includes 30,324 high-confidence annotations across 116 species. Files in this collection Audio recordings can be accessed by downloading and extracting the “soundscape_data.zip” file. Soundscape recording filenames contain a sequential file ID, site ID, recording date, and timestamp in IST. As an example, the file “WGH_0106_001_20181201_072000.flac” has sequential ID 0160 and was recorded at site S01 on Dec 1st, 2018 at 07:20:00 IST. Audio files are between 4 and 5 minutes long. We also provide 4,800 additional soundscape recordings, randomly selected from the same deployment that have not been annotated but can be used for self- or unsupervised training. These files can be accessed by downloading and extracting the "unlabeled_soundscape_data.zip" file and follow the naming scheme above. Ground truth annotations are listed in “annotations.csv” where each line specifies the corresponding filename, start and end time in seconds, low and high frequency in Kilohertz, and an eBird species code. These species codes can be assigned to the scientific and common name of a species with the “species.csv” file. The approximate recording location with Universal Transverse Mercator (UTM) coordinates and other metadata can be found in the “recording_location.csv” file. Acknowledgements Compiling this extensive dataset was a major undertaking, and we are very thankful to the domain experts who helped to collect and manually annotate the data for this collection. Specifically, we would like to thank Vijay Kumar and Krishna Kumar, members of the Kadar indigenous forest community, whose knowledge of the landscape immensely helped with fieldwork and the deployment of audio recorders in the Anamalai hills. We thank the plantation management of Tata Coffee Ltd., Parry Agro Industries Limited and Tea Estates India Limited for permission to work in their estates. Thanks to the Tamil Nadu Forest Department for providing the necessary permissions to Dr. V.V. Robin (permit no: WL(A)/22030/2019) to access protected areas, which enabled us to carry out this work. We thank Dr Divya Mudappa and Dr T.R. Shankar Raman at the Nature Conservation Foundation (N.C.F.) for all their technical and logistical support. We would also like to acknowledge our funding sources: Funding from the National Geographic Society, Explorers Club, Rufford Foundation and the Edward W. Rose Postdoctoral Fellowship was instrumental in completing this work.