Audio environment classification for hearing aids.

摘要

This thesis examines background classification systems for hearing aids. First, the K-nearest neighbours classifiers (KNN), hidden Markov models (HMM), multi-layer perceptrons (MLP) and MLPs with windowed input (WMLP) are assessed for functionality. K-means clustering and self-organizing maps (SOMs) are used to find appropriate classes. The classes selected are in-car, traffic, birds, water washing, water running, office, restaurant, shopping and music. Feature selection is then performed on a large candidate set from literature, using sequential forward floating search (SFFS). The classifiers are then tested using the classes and feature vector selected, and using a simple feature vector from literature. Using features selection improves the results for the KNN, but not the other classifiers. The KNN gives an average accuracy of 77.6% and a best-run accuracy of 84.4%. The WMLP gives an average accuracy of 65.6% and a best-run accuracy of 80.0%, but requires less memory and processing than the KNN.