GTM-UVigo systems for Albayzin 2014 Search on Speech Evaluation

This paper describes the systems developed by the GTMUVigo team for the Albayzin 2014 Search on Speech evaluation. The primary system for the spoken term detection task consisted on the fusion of two different large vocabulary continuous speech recognition systems that differed in almost all their components: front-end, acoustic modelling, decoder and keyword search approach. An isolate word recognition system was fused with the two aforementioned speech recognition systems for the keyword spotting task. For the query by example spoken term detection task, a fusion of three systems was presented: one of them followed one of the aforementioned continuous speech recognition approaches, with the difference that in this case it was necessary to obtain a transcription of the queries; the other two systems performed a dynamic time warping search, being the use of fingerprints as feature vectors the main novelty of the presented approach.