Dipl.Ing. Mirko Hannemann
| Černocký, J., Szőke, I., Hannemann, M., Kombrink, S., Fapšo, M.: Hybrid Word-Subword Speech Recognition - a Powerful Tool to Search in Speech, Proceedings of 21st International Conference Radioelektronika 2011, Brno, CZ, UREL FEKT VUT, 2011, p. 25-25, ISBN 978-1-61284-322-3 | | Publication language: | english |
|---|
| Original title: | Hybrid Word-Subword Speech Recognition - a Powerful Tool to Search in Speech |
|---|
| Title (cs): | Hybridní slovní a podslovní rozpoznávání řeči - výkonný nástroj pro vyhledávání v řeči |
|---|
| Pages: | 25-25 |
|---|
| Book: | Proceedings of 21st International Conference Radioelektronika 2011 |
|---|
| Conference: | Radioelektronika 2011, 21st International Conference |
|---|
| Place: | Brno, CZ |
|---|
| Year: | 2011 |
|---|
| ISBN: | 978-1-61284-322-3 |
|---|
| Publisher: | Department of Radioelectronics FEEC BUT |
|---|
| URL: | http://www.fit.vutbr.cz/research/groups/speech/publi/2011/cernocky_radioelektronika2011_150_invited.pdf [PDF] |
|---|
| Abstract |
|---|
| Main-stream systems for searching information in speech are based on Large Vocabulary Continuous Speech
Recognizer (LVCSR) with fixed vocabulary. The keywords or key-phrases are subsequently searched in its output. These
systems have severe problems with Out of Vocabulary (OOV) words, that are common when one changes the domain (for
example from standard to medical), speaker (normal versus highly educated), or even date (new words appearing in TV
news). This talk will present our work in designing hybrid word-subword recognition systems, that have a combined
recognition network. Under normal circumstances, they output standard word strings, while they are allowed to switch to
subword description for unknown inputs. Such systems are good not only for detecting OOVs, but also subsequent steps
leading to their exploitation. Under the EC-sponsored DIRAC project, we have investigated analysis of detected OOVs,
conversion to standard word-form, and finding links to in-vocabulary words and other OOVs. The results will be
demonstrated on real speech data from popular TED lectures. |
|