When: Sunday 1:15pm to 3:15pm
Organiser: Louis Ranjard, ANU, Canberra
Passive bioacoustic monitoring is revolutionising ecology by allowing the rapid and long-term collection of very large amounts of data in the field. A challenge now is to make sense of this data while ensuring reproducibility of the statistical analyses. In the recent years, several machine learning approaches have been developed to allow researchers to analyse long field recordings. In this session, I will demonstrate the use of MatlabHTK, a software interface to extract signal of interest from long recordings using hidden Markov models (HMMs). MatlabHTK is built upon approaches that were originally developed for speech processing and have been adapted to bioacoustic analysis. The analysis is performed in two steps: first, HMMs are built on training data that has been manually annotated and, second, the long recordings are processed. The different
parameters that can be tuned to encode the signal will be discussed. I will then show some possible subsequent analyses that can be performed on the processed recordings. In particular, I will present the use of DTWave_cluster to perform unsupervised clustering of the extracted signals of interest, e.g. vocalizations, using an artificial neural network approach. This tool intents to take into account uncertainty in the definition of the underlying clusters and, thus, permits robust statistical analysis to uncover patterns of variation in the data.