Discriminative Filtering ─ a Non-ASR Approach to Keyword Spotting
Prof. Prasanta Kumar Ghosh
Indian Institute of Science, India
Discriminative training is a tool used to train a model to maximally separate favorable from competing classes. Discriminative training techniques, although popularly used in Automatic Speech Recognition (ASR) systems now-a-days, have not yet been widely applied to Keyword Spotting (KWS) applications. Amongst Non-ASR based KWS systems, Point Process Model (PPM) based KWS has become a popular research topic in recent days. In this presentation, we will talk about an extension of PPM-based KWS to a discriminative paradigm and propose a PPM training algorithm for a low training resource situation. We will also talk about discriminative 2-D filtering on phoneme posteriorgrams to maximally separate the filter response at the keyword and competing keyword locations – Level Discriminative Optimal (LDO) filters, being one example of such a discriminative filter which is trained to assign the filter response at keyword and non-keyword locations in speech to two distinct levels.
Research keywords : Human speech communication science, Speech science application