Acoustic feature similarity between utterances has been shown to be very helpful for spoken term detection using pseudo-relevance feedback (PRF) and graph-based re-ranking. Both cases are based on the concept that utterances similar to those utterances with higher relevance scores in acoustic features should have higher scores, while graph-based re-ranking further considers the similarity structure between many utterances globally with a graph. In this paper, we extend these approaches to consider acoustic feature similarity between utterances over both word and subword lattices, and offer a complete formulation for the general problem of open vocabulary retrieval of spoken content with shorter or longer queries. All these are verified by significant improvements in preliminary experiments with both in vocabulary (IV) and OOV queries.

Index Terms: Spoken Content Retrieval, Pseudo-relevance Feedback, Random Walk