This study focuses on recognizing bird species from their voices, which are frequently seen in Aras River Bird Sanctuary of Iğdır. For this purpose, deep learning methods were used. Acoustic monitoring is carried out to examine and analyze biological diversity. Passive acoustic listeners/recorders are used for this work. In general, various analyzes are performed on the raw sound recordings collected with these recording devices. In this study, raw sound recordings obtained from birds were processed with the methods developed by us, and then bird species were classified with deep learning architectures. Classifications were carried out on 22 bird species that are frequently seen in Aras Bird Sanctuary. Audio recordings were made into 10-second clips and then converted into one-second log mel spectrograms. Convolutional Neural Networks (CNN) and Long Short-Term Memory Neural Networks (LSTM), which are deep learning architectures, were used as classification methods. In addition to these two models, the Transfer Learning method was also used. Highlevel feature vectors of sounds were extracted with VGGish and YAMNet models, which are pre-trained convolutional neural networks, used for transfer learning. These extracted vectors formed the input layers of the classifiers. Accuracy rates and F1 scores of four different architectures were found through experiments. Accordingly, the highest accuracy rate (acc) and F1 score were obtained with the classifier using the VGGish model with 94.2% and 92.8%, respectively.

No abstract

This study presents a comparison of four different machine learning algorithms for sentiment analysis on a small subset of the AR-P (Amazon Reviews - Polarity) dataset. The algorithms evaluated are multilayer perceptron (MLP), Naive Bayes, Decision Tree, and Transformer architectures. The results show that the Transformer-based DistilBERT model performed the best with an accuracy rate of 96.10%, while MLP had a better performance than the other remaining methods. Confusion matrices and ROC curves are provided to illustrate the results, and a comparison with previous studies is presented. The study concludes that the results can serve as a basis for future work, such as using larger datasets or comparing the performance of algorithms on different tasks. Overall, this study provides insights into the use of traditional machine learning and modern deep learning methods for sentiment analysis and their potential applications in real-world scenarios.