Abstract:
The motivation behind this thesis is to remove hazardous sound sources from a work environment, whilst leaving some sounds that can be used for safety cues; particularly focusing on construction and factory sites. This concept is of interest because hearing loss from the working environments is a big issue; therefore prevention needs to be put in place. The problem and foundation of this thesis is separating and cancelling unwanted sound. To approach this problem, Independent component analysis (ICA) is used, and three different ICA algorithms; kurtosis, negentropy and infomax estimation are investigated. These algorithms can separate source signals by using either gradient ascent or fast ICA. Once the signals have been separated, they are inverted and summed into the signal mixtures to cancel the unwanted sound source. The algorithms need to be tailored based on the problem stated above; therefore the speed of convergence needs to be increased, this was achieved by implementing head tracking. Another improvement is to add the ability to cancel changing sound sources. To achieve this, ICA requires a model of the source signal distribution that is to be separated. Two methods were developed that can model a distribution on-line and can be directly used in the ICA algorithms. It was found that with the aid of head tracking and custom models, ICA was tailored to separating and cancelling unwanted sound. Overall there was an increase in convergence speed and quality of the extracted source signal. Based on the results, we can separate and cancel unwanted sound sources in working environments that this thesis specifically addresses; which has practical implications for improving sound protection equipment.
