Semi-supervised classification of acoustic signals

The project focuses on semi supervised acoustic signals classification problem, in which we cluster elements into different groups. The raw data consists of bats echolocation recordings from which pulses were extracted, each pulse related to a single bat.
The project goal is to recognize the generator of the pulses based on the pulse waveform only.
We extract features from these pulses to generate an input matrix for the clustering process. The main clustering steps are dimensionality reduction and semi supervised clustering.
We examine the performance of several algorithms for dimensionality reduction (PCA, Kernel PCA, Spectral Analysis and tSNE) as well as different clustering methods (K-Means, GMM).
Using each of these methods, applied on the pulses’ amplitude modulation as a feature, the classifier we build achieves high accuracy (more than 90% in average) in classifying the pulses to their clusters.

Semi-supervised classification of acoustic signals
Semi-supervised classification of acoustic signals
Collaboration:

Rafael