Background

This project was inspired by a bird sound detetection challenge. An algorithm based on logistic models was developed to determine where there are birds calling in a specific sound recording. The performance of this algorithms was further compared to a CNNs model. Detaild report can be found here. Due to the space limitation, only part of the data were used here to demonstrate two algorithms. Full dataset can be found in the official site of the challenge.

If applying to the new dataset is of interest, simply replace the data in the soundData folder!

Preparation

• Clone or download the Bird-Sound-Detection-2019 repo to the local computer
• Install R version 3.6.0
• Install required packages.
  • For packages from CRAN..

  install.packages("Name_of_Package")

  • For packages from Bioconductor, such as EBImage, please follow the instruction.

  install.packages("BiocManager")
  BiocManager::install("Name_of_Package"), 

Comparison between logistic model and CNNs

Here is the result based on randomly dividing the full dataset: 12552 (80%) recordings for training and 3138 (20%) recordings for making prediction, i.e., testing.

Processing time for training (hr)	Processing time for prediction (hr)	Accuracy (%)
1:09	0:17	76
6:48	1:40	84

Have fun!