Phd student Ewha Womans University, Republic of Korea
Automated call detection is a challenging aspect of bioacoustic research, as it allows easier acoustic monitoring. While various methods have been proposed, many are time-consuming and require large amounts of training data. R-ribbit, originally developed for frog call detection, provides a relatively uncomplicated method for detection and does not require an extensive amount of training data. However, its use has been limited to frog calls, and it is unclear whether it can be effective for bird call detection. In this study, we aim to determine the optimal parameters for R-ribbit to detect common crow calls, and evaluate its effectiveness as a tool for acoustic monitoring of birds. As our target species, the common crow was selected, as it has a lower repetition rate and longer duration compared to frog calls used for R-ribbit. We collected recordings of the common crow calls and identified the input parameters that derived the most accurate detection results (Spectrogram window length: 4096 samples; window overlap: 0; signal band: 600-3000Hz; pulse rate: 0.3-2 pulse/s; analysis window length 4s). To achieve this, We used the logistic regression analysis to train the model on the common crow recordings. We created a confusion matrix to test its performance. Our results show that our model was significantly appropriate (z =3.06, p=0.002*). R-ribbit achieved high accuracy (92.9%) for the common crow call detection. However, its precision (2.6%) and recall scores (14.2%) were low due to a lack of true positives in the test dataset. Our findings suggest that R-ribbit has the potential to be an effective tool for bird call detection, but further validation with larger datasets is needed to measure its balanced accuracy for the biased dataset.
