Machine listening in combination with microphone arrays for noise source localization and identification

In a recent project, a large microphone array system has been created to localize and quantify noise sources in an Intensive Care Unit (ICU). In the current state, the output of the system is the location and level of the most dominant noise sources, which is also presented in real-time to the nursing staff. However, both staff as well as patients have expressed the need for information about the types of noise sources. This additional source identification can also help to find means of reducing the overall noise level in the ICU. To accomplish the source identification, the approach of machine listening with a deep neural network is chosen. A feed-forward pattern recognition network is considered in this work. However, it is not clear which types of features are best suited for the given application. This contribution thus examines the problem from a practical point of view, comparing different features including those related to sound perception, such as specific loudness, Mel-frequency cepstral coefficients, as well as the output of a gamma-tone filter bank. Additionally, the concept of time-delay networks is tested to see whether a better classification of the signals can be achieved by including their time history.