Pervasive Sound Sensing: A Weakly Supervised Training Approach
Refereed Original Article
Modern smartphones present an ideal device for pervasive sensing of human behavior. Microphones have the potential to reveal key information about a person’s behavior. However, they have been utilized to a significantly lesser extent than other smartphone sensors in the context of human behav- ior sensing. We postulate that, in order for microphones to be useful in behavior sensing applications, the analysis tech- niques must be flexible and allow easy modification of the types of sounds to be sensed. A simplification of the training data collection process could allow a more flexible sound clas- sification framework. We hypothesize that detailed training, a prerequisite for the majority of sound sensing techniques, is not necessary and that a significantly less detailed and time consuming data collection process can be carried out, allow- ing even a nonexpert to conduct the collection, labeling, and training process. To test this hypothesis, we implement a diverse density-based multiple instance learning framework, to identify a target sound, and a bag trimming algorithm, which, using the target sound, automatically segments weakly labeled sound clips to construct an accurate training set. Experiments reveal that our hypothesis is a valid one and results show that classi- fiers, trained using the automatically segmented training sets, were able to accurately classify unseen sound samples with accuracies comparable to supervised classifiers, achieving an average F -measure of 0.969 and 0.87 for two weakly supervised datasets.
Digital Object Identifer (DOI):
Cybernetics, IEEE Transactions on (Volume:PP , Issue: 99 )
National University of Ireland, Dublin (UCD)
Open access repository: