@InProceedings{Elmers2022_1160,
<br/>   author = {Mikey Elmers},
<br/>   booktitle = {Studientexte zur Sprachkommunikation: Elektronische Sprachsignalverarbeitung 2022},
<br/>   title = {Comparing detection methods for pause-internal particles},
<br/>   year = {2022},
<br/>   editor = {Oliver Niebuhr and Malin Svensson Lundmark and Heather Weston},
<br/>   month = mar,
<br/>   pages = {204--211},
<br/>   publisher = {TUDpress, Dresden},
<br/>   abstract  = {This study investigates different machine learning architectures for classifying
pause-internal phonetic particles (PINTs), such as filler particles (FPs),
breath noises complementary to silences, and tongue clicks. Many of these PINTs
co-occur, and by modeling them simultaneously, the aim is to improve the classification
accuracy for the surrounding PINTs as well. An annotated subset from a
German spontaneous speech corpus was used for modeling. Mel-frequency cepstral
coefficients were used as inputs to model PINTs with three kinds of neural
networks: a general neural network, a convolutional neural network, and a recurrent
neural network. The models used the same hyperparameters, number of layers,
and number of neurons for those layers, so that the focus was put onto the model
architecture. The recurrent neural network was expected to perform the best since
it is able to capture temporal information; however, all models performed similarly.
The models performed best at classifying silent segments, followed by inhalations
and exhalations. However, all models failed to accurately classify FPs and clicks,
indicating that modeling PINTs simultaneously doesn’t always improve accuracy
for surrounding PINTs. These findings suggest that accurate classification is more
dependent on annotation quantity and quality than model architecture. The main
contributions of this paper are the classification of multiple PINTs simultaneously,
and the improvement of PINTs classification for the German language.},
<br/>   isbn = {978-3-95908-548-9},
<br/>   issn = {0940-6832},
<br/>   keywords = {Signal Processing & Comprehension},
<br/>   url = {https://www.essv.de/pdf/2022_204_211.pdf},
<br/>}