AUREAS: a tool for recognition of anuran vocalizations
Implementing techniques that facilitate and automatize the species monitoring is a task that can be carried out in several ways. One of them is bioacoustics analysis, which focuses on analyzing the soundscapes of a large number of recordings. One of the bioacoustics monitoring methods is species call recognition, which can be tackled commonly using four stages: cleaning, segmentation, feature extraction, and classification. Depending on the methodology several stages can be omitted and additional stages can also be considered, e.g., feature selection. In order to propose a monitoring system, the research groups SISTEMIC and Grupo Herpetologico de Antioquia (GHA) of the Universidad de Antioquia have implemented techniques focused on anuran calls recognition, specifically fuzzy clustering algorithms. In the segmentation stage, only the anuran vocalization segments were extracted from the spectrogram, ignoring the remaining frequencies. This procedure allows removing noise that can be found at other frequencies. Then, for the feature extraction stage, descriptors based on the MFCC (Mel Frequency Cepstral Coefficients) were computed, but with the difference that Mel scale was removed and the coefficients were computed only from the segmentation of the call. For the classification stage, an unsupervised algorithm called LAMDA (Learning Algorithm for Multivariate Data Analysis) was used to classify the segmented calls. LAMDA allows creating new classes that were not included in the learning process but were presented in the recognition step. The complete methodology was used to create the software AUREAS, which was tested using a database of 7 anuran species (1712 segments, including segments where there are no calls) from the northern Andes of Colombia. This software can identify these anuran calls with a F1-score of 0.88, allowing us to obtain activity patterns of the anuran species, which is a useful tool to monitor these species at the time. Now, the software is being modified to be able to recognize avian calls that have generally more inter-species variation. Therefore, more features based on other approaches (Wavelet-based, linear predictive codes, perceptual linear prediction) were included in order to abstract different attributes of the calls. To identify the most informative features, a feature selection stage was included, which used wrapper and filter strategies. In this conference, the methods implemented in the software and the different study cases to identify anuran and birds calls will be presented. We will illustrate the main challenges that are still required to solve issues in segmentation, feature extraction, and classification stages.