A Novel Zero-Resource Spoken Term Detection Using Affinity Kernel Propagation with Acoustic Feature Map

Spoken term detection (STD) without linguistic clues is challenging for retrieval tasks. Despite numerous studies to overcome the challenges, there is a scope for improvement. Dynamic time warping based techniques were extensively employed to accomplish the STD task in the absence of linguistic resources. A drawback of this approach is handling the speaker, language, acoustic and spoken query variabilities that exist in natural speech. Our approach introduces a novel acoustic feature representation adjoined with affinity kernel propagation to overcome the challenges. At first, the Self Organising Map based feature vector representation was employed to overcome the speaker variability issues. In the next stage, introducing the affinity kernel propagation approach captures the best alignment between the spoken query and the utterances in the similarity-matching task without constraining the nature of the query. By introducing the acoustic feature mapping and similarity-matching through affinity kernel propagation, a 6% performance gain of Maximum Term Weigh Value and a 5% reduction in the cross-entropy cost were achieved during the evaluation with QUESST-14 speech corpus across multiple languages.