Title: | Detection of overlapping speech using a convolutional neural network: first experiments |
Authors: | Kunešová, Marie |
Citation: | RENDL, Jan ed. Studentská vědecká konference: magisterské a doktorské studijní programy, sborník rozšířených abstraktů, květen 2018, Plzeň. Plzeň: Západočeská univerzita v Plzni, 2019, s. 56-57. ISBN 978-80-261-0790-3. |
Issue Date: | 2018 |
Publisher: | Západočeská univerzita v Plzni |
Document type: | konferenční příspěvek conferenceObject |
URI: | http://hdl.handle.net/11025/29824 http://svk.fav.zcu.cz/download/sbornik_svk_2018.pdf |
ISBN: | 978-80-261-0790-3 |
Keywords: | konvoluční neuronové sítě;počítačové zpracování řeči;překrývající se řeč |
Keywords in different language: | convolution neural networks;computer speech processing;overlapping speech |
Abstract in different language: | Many speech processing applications, such as speaker diarization and speech recognition, have problems with overlapping speech, i.e. intervals in which multiple speakers are talking simultaneously. This happens particularly often in spontaneous conversations, where speakers may regularly interrupt each other or interject short utterances while the original speaker keeps talking. Detecting such occurrences can help improve the performance of the impacted systems.However, this is still an actively researched task, which has not yet been fully solved. In this work, I describe my initial experiments in using a convolutional neural network (CNN) to detect overlapping speech in an artificial dataset created for the purpose. |
Rights: | © Západočeská univerzita v Plzni |
Appears in Collections: | Studentská vědecká konference 2018-magisterské a doktorské studijní programy Studentská vědecká konference 2018-magisterské a doktorské studijní programy |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
Kunesova.pdf | Plný text | 382,8 kB | Adobe PDF | View/Open |
Please use this identifier to cite or link to this item:
http://hdl.handle.net/11025/29824
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.