Audio signal one-class classification over the years

The first time I encountered the one-class classification of audio di was back in 2009. Back then, I wasn’t too keen to try machine learning approaches to the problem, so I stuck to what was familiar to me: convolution. This problem was different from voice recognition because humans can produce a wider range of sounds than the ones I was trying to detect, namely sounds produced by certain insect species.

Some alternatives were presented to me then:

• Window functions (e.g. Hamming window)
• Matched filter
• Wiener filter
• Dynamic time warping
• Matlab’s MA Toolbox audio similarity measure
• Support vector machine
• Blind source separation (e.g. Independent component analysis)
• Hidden Markov Model
• Naive Bayes Classifier
• Autocorrelation via recursive least squares filter

Then in 2022 I finally came accross a solution involving neural networks: recurrent variants (LSTM/GRU) that can be paired with one or two convolutional layers (CNN), with the CNN layer operating on a few frames of FFT-transformed audio.