Comparative Analysis of MFCC and Convolutional Autoencoders for the Detection of Dendropsophus Sarayacuensis

Tuesta D.P.A.T.; Teagua K.J.R.; Rios K.A.T.; Satalaya A.M.N.; Vigo R.C. · International Conference on Electrical, Computer, and Energy Technologies, ICECET 2026 (2026) · DOI: 10.1109/icecet65726.2026.11632896

Resumen

Passive acoustic monitoring (PAM) is a powerful tool for assessing biodiversity; however, the automated analysis of recordings in acoustic complex environments, such as the Amazon rainforest, remains a difficult task. This paper presents a comparative study of the performance of feature extractors for the automatic detection of vocalizations by Dendropsophus sarayacuensis, a frog native to the Peruvian Amazon and a bioindicator of tropical ecosystems. Thus, Mel-frequency cepstral coefficients, convolutional autoencoders, and variational autoencoders were evaluated in 13D and 128D configurations, combined with SVM, Random Forest, and KNN, resulting in 15 combinations evaluated using 5-fold cross validation on 2,669 0.33-second segments collected in the Alpahuayo-Mishana National Reserve, Loreto, Peru. Analysis of the results showed that the highest accuracy was achieved by CAE-13D with cosine KNN (94.78%), outperforming the best MFCC result of 94.39% with linear SVM. Which means that the results demonstrate that performance depends on the interaction between the feature extractor and the classifier: MFCC generates linearly separable spaces favorable for SVM, while CAE produces compact representations that KNN exploits efficiently. It can be concluded that the variational regularization of the VAE exhibited a classifier-dependent effect, improving SVM performance while simultaneously penalizing KNN, so these results contribute to the design of acoustic detection methods for monitoring Amazonian biodiversity.

Instituciones: Universidad Nacional de la Amazonía Peruana