Glass Frog Calls

This dataset contains 5296 audio files in .WAV format, corresponding to calls of two glass frog species: Hyalinobatrachium fleischmanni (Hf) and Espadarana prosoblepon (Ep), recorded under controlled laboratory conditions. We further increased the dataset size using noise injection-based data augmentation in one species (Ep), and artificially shifting call frequency in both species. The dataset contains four labelled classes: Hf (1250 calls), Hf-shifted (1250 calls), Ep (1398 calls) and Ep-shifted (1398 calls); each class is in a separate folder.

Noise injected files, with white and pink noise, were generated using noise factors of 0.010 and 0.020, respectively, using the random and powerlaw_psd_gaussian functions from numpy (v1.26.4) and colorednoise (v2.2.0) Python libraries.

Frequency shifting placed calls in upper or lower parts of spectrum, incorporating additional variability and was used to create two new classes. Hf call frequency was increased by a factor of +4 semitones (4/12), while Ep call frequency was decreased by the same factor (-4), using librosa (v0.10.2) in Python (effects.pitch_shift). Modified calls (‘Hf-shifted’ and ‘Ep-shifted’, respectively) more closely resemble the calls of the other species in terms of their frequency band.

Derived audio files partially share file names: '1104-18211-ROI1_pN.WAV' is the 'pink noise' version of '1104-18211-ROI1.WAV', etc.

The audio collection is accompanied by a data table (.csv) with no missing or null values, and consists of 4 columns: “File_name”; “Data_Augmentation”, whether the file had noise injection or not, and its type (white or pink); “Frequency_Shift”, whether the audio frequency was artificially shifted or not; and “Class”, which includes the respective labels. The dataset is suitable for machine learning tasks, audio signal processing and statistical analysis.