Beyond Posteriorgram: Bottleneck Features for Keyword Search

Template matching approaches have been proposed as an alternative to large vocabulary continuous speech recognition (LVCSR) based systems for keyword search. These have been shown to have no performance discrepancy between terms in the training vocabulary and out of vocabulary (OOV) terms. Those methods have often relied on the use of posteriorgram as the features for the search. In this paper, we propose the use of bottleneck features instead of posteriorgram because of their potential for cross-lingual transfer learning. We show the feasibility of bottleneck features for template-matching based keyword search using by learning different representations for the features including a mixture of Gaussians and a representation based on a joint distance metric learning framework.