Speech utterance categorisation given one training utterance per category

In this paper, we address the categorisation of speech utterances within the scenario of technical support automated agents given only one labelled utterance per category . The categorisation algorithm maps input utterances into bag-of-word vectors and then applies feature extraction based on soft word clustering. We analyse two feature extraction schemes: pole-based overlapping clustering (PoBOC) and a combination of PoBOC with Fuzzy c-medoids. For the categorisation at the utterance level, we use the Nearest Neighbour (NN) approach. Finally, we evaluate the proposed methods on a test corpus with more than 3000 utterances recorded in a commercial dialog system.