A natural acoustic indicator of animal welfare is the appearance (or absence) of coughing in the animal habitat. A sound-database of 5319 individual sounds including 2034 coughs was collected on 6 healthy animals containing both animal vocalisations and background-noises. Each of the test-animals was repeatedly placed in a laboratory installation where coughing was induced by nebulization of citric acid. A 2-class classification into 'cough' or 'other' was performed by application of a distance function to a fast fourier spectral sound analysis. This resulted in a positive cough-recognition of 92%. For the whole sound-database however there was a mis-classification of 21%. As spectral information up to 10000 Hz is available an improved overall classification on the same database is obtained by applying the distance function to 9 frequency-ranges and combining the achieved distance-values in fuzzy rules. For each frequency-range clustering-threshold is determined by fuzzy C-means clustering.
展开▼
