A novel approach for bird sound classification with cross correlation by denoising with complementary ensemble empirical mode decomposition using B-spline and LSTM features

Abstract

This research covers the application of modern signal processing and machine learning techniques for the purpose of denoising and accurately classifying bird sounds. In the study, the signal is separated into smaller components by using the CEEMD (Complete Ensemble Empirical Mode Decomposition) method to clean the background noise of bird sounds. Then, B-Spline functions and LSTM (Long Short-Term Memory) network are used for feature extraction on these cleaned signals. These two methods are effective in capturing important trends and hidden information in the sound signals over time, and a richer feature vector is created. The most important stage of the research is the application of cross-correlation to this feature vector. Cross-correlation provides a powerful analysis tool in terms of timing and pattern detection by analyzing the time delays and similarities between two signals. This process played a major role in determining the similarities between sound signals and increased the classification performance. After feature extraction and cross-correlation, classification is performed using different machine learning algorithms. In general, when cross-correlation is applied, the performance of all algorithms is significantly increased and especially with the C-SVM algorithm, 98.32% accuracy and 98.62% F1 score are obtained. These results show that cross-correlation is a powerful tool in the classification of sound signals and high accuracy rates are achieved when used together with methods such as CEEMD, B-Spline and LSTM. The results of this study show that modern signal processing techniques are effective in the analysis of complex sound signals such as bird sounds and cross-correlation is a critical step in improving the classification performance.