Automatic recognition of ingestive sounds of cattle based on hidden Markov models

Information about ingestive events like chewing and biting is useful for estimation of intake and monitoring of grazing behaviour. We present an automatic tool to decode ingestive sounds of cattle into ingestive events. Ingestive sounds can be recorded easily and without alteration of normal grazing behaviour by placing a microphone on the forehead of the animal. However, recorded sound need to be decoded automatically for the method to be of practical use. Hidden Markov models have been successfully used to segment and classify acoustic signals. In this work we extend the use of hidden Markov models to recognise ingestive sounds of cattle. We present new findings about the spectral content of the acoustic signals and a novel language model for the recognizer. Three types of ingestive events (bites, chews and chewbites) by cows grazing tall (24.5 ± 3.8 cm) or short (11.6 ± 1.9 cm) alfalfa or fescue were successfully recognised. Recognition rates were 84% for tall alfalfa, 65% for short alfalfa, 85% for tall fescue and 84% for short fescue. These levels of correct classification are suitable for quantification of grazing behaviour.

► High recognition rates have been obtained with two different pastures and heights. ► A new language model is provided to increase the performance of the model in the long-term. ► Recognizer discriminates not only bites but also chews and the combined chew–bite events. ► The most relevant information to discriminate the chew, bite and chew–bite events was found below 500 Hz.