Schneider, Daniel: Holistic Vocabulary Independent Spoken Term Detection. - Bonn, 2012. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5n-28581
@phdthesis{handle:20.500.11811/5314,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5n-28581,
author = {{Daniel Schneider}},
title = {Holistic Vocabulary Independent Spoken Term Detection},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2012,
month = jun,

note = {Within this thesis, we aim at designing a loosely coupled holistic system for Spoken Term Detection (STD) on heterogeneous German broadcast data in selected application scenarios. Starting from STD on the 1-best output of a word-based speech recognizer, we study the performance of several subword units for vocabulary independent STD on a linguistically and acoustically challenging German corpus. We explore the typical error sources in subword STD, and find that they differ from the error sources in word-based speech search. We select, extend and combine a set of state-of-the-art methods for error compensation in STD in order to explicitly merge the corresponding STD error spaces through anchor-based approximate lattice retrieval. Novel methods for STD result verification are proposed in order to increase retrieval precision by exploiting external knowledge at search time. Error-compensating methods for STD typically suffer from high response times on large scale databases, and we propose scalable approaches suitable for large corpora. Highest STD accuracy is obtained by combining anchor-based approximate retrieval from both syllable lattice ASR and syllabified word ASR into a hybrid STD system, and pruning the result list using external knowledge with hybrid contextual and anti-query verification.},
url = {https://hdl.handle.net/20.500.11811/5314}
}

The following license files are associated with this item:

InCopyright