Towards Vocabulary-Independent Speech Indexing For Large-Scale Repositories

The Out-Of-Vocabulary problem remains a challenge for word-lattice-based speech indexing. Sub-word-based approaches address this problem effectively for small-scale tasks, but suffer from poor precisions on large-scale databases due to lack of strong language model constraints. We propose a method for searching OOV queries with large-scale databases in two steps. First, result candidates are extracted from a sub-word-based system, ensuring a high recall. The candidates are then refined by word-lattice rescoring aiming at a high precision. Experiments on a 160-hours lecture set show that the proposed approach achieves a relative improvement of 8.7% over the sub-word-based baseline, and 19.7% for only single-word queries.