Classifying Bird Songs Based on Chroma and Spectrogram Feature Extraction

An increasingly important computational tool in ecology, conservation monitoring, and vocal communication studies is the automatic species classification of birds based solely on their auditory characteristics. Birds share all the characteristics of an animal because they share a common ancestor with all other animals on the planet. Birds are considered animals. Birds are vertebrate animals, improving classification accuracy and making sure it can operate at large data scales are essential for making it practical. Numerous methods rely on spectrogram-type data for acoustic measures, such as the features of the Mel-frequency cepstral coefficient (MFCC), a manually created summary of spectral data. On the other hand, current machine learning research has shown that automatically learned features from data can frequently outperform feature transforms created by hand. The classification of bird song is crucial for keeping track of the bird population in various settings. This paper tries to offer new acoustic feature for music analysis and speech like Chroma. Chroma does so with reference to the birdsong dataset, which has a complex and varied audio background. It is combined and spliced with the widely used component of birdsong that is Log-Mel Spectrogram (LM) to enhance the representational power of a single feature; In addition, Support Vector Machine, Random Forest and CNN are proposed as a pretrained classifier to increase the sensitivity to the information of birdsong that varies over time considering the fact that birdsong has dynamic and continuous changes over the time. To assess how Chroma and LM were fused to maximise the information of the audio included in the recordings of birdsong, we used specific bird audio data from the Kaggle website. The findings of the experiment indicate that the best result, a 99.58% accuracy rate, was obtained using the LM-C feature combination and CNN model.