Analyzing Cricket Songs with Machine Learning
Loading...
Date
2022-04-14
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Given the recordings of cricket songs, we aimed to apply machine learning and audio signal processing techniques to (1) discover intrageneric and intraspecies relationships in the songs and (2) create models that could classify the songs into their correct genus and species. First, we took out noise in the audio files using a high pass filter. Then, we represented the cricket songs in three different forms: mel spectrograms, mel frequency cepstrum coefficients, and magnitude power spectrums. We achieved our first objective by reducing the dimensionality of the three extracted audio features to visualize how the cricket songs clustered in 2D space. We found that cricket songs belonging to the same genus are generally similar, which gave us hope that constructing a high accuracy genus classification model would be possible. We were not able to conclude that cricket songs belonging to the same species are similar because the dataset used did not have enough audio files per species. As a result of our initial findings, we constructed genus classification models using shallow convolutional neural network architectures and mel spectrograms as input. Because there was only a small number of cricket song files available for training, we reduced our models’ scopes to only classify inputs into 5 genera. Rather than extracting a single mel spectrogram per each available audio file, we extracted multiple 3-second mel spectrograms per each available audio file and used this set of mel spectrograms as our training set. We found that the more mel spectrograms extracted per available audio file, the higher the model’s accuracy became. Our highest-performing genus classification model has a validation loss of 0.1637 and a validation accuracy of 94.30%. Later in the research, we obtained a larger dataset that could be used for species classification. As such, we constructed species classification models using the same approaches taken for the genus classification models. These models classified inputs into 9 different species. Our highest-performing species classification model has a validation loss of 0.2849 and a validation accuracy of 92.28%. The high accuracy of our genus and species classification models confirmed that it would be possible to classify cricket songs using machine learning techniques. This finding is pivotal for the entomology field as there has not been much documented research regarding insect song classification. Of course, our classification models were quite limited in scope; however, we believe that with more data and a deeper model, more generalized insect song classification models are possible. The fact that we were able to employ simple yet effective techniques to discover insights in relatively small cricket song datasets should encourage entomologists that the application of machine learning to cricket songs is still worthwhile even if the field has a lack of publicly available data.
Description
Keywords
Machine Learning, Crickets, Cricket songs, Entomology